git/pack-bitmap-write.c

701 lines
17 KiB
C
Raw Normal View History

pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
#include "cache.h"
#include "object-store.h"
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
#include "commit.h"
#include "tag.h"
#include "diff.h"
#include "revision.h"
#include "list-objects.h"
#include "progress.h"
#include "pack-revindex.h"
#include "pack.h"
#include "pack-bitmap.h"
#include "hash-lookup.h"
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
#include "pack-objects.h"
#include "commit-reach.h"
#include "prio-queue.h"
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
struct bitmapped_commit {
struct commit *commit;
struct ewah_bitmap *bitmap;
struct ewah_bitmap *write_as;
int flags;
int xor_offset;
uint32_t commit_pos;
};
struct bitmap_writer {
struct ewah_bitmap *commits;
struct ewah_bitmap *trees;
struct ewah_bitmap *blobs;
struct ewah_bitmap *tags;
kh_oid_map_t *bitmaps;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
struct packing_data *to_pack;
struct bitmapped_commit *selected;
unsigned int selected_nr, selected_alloc;
struct progress *progress;
int show_progress;
unsigned char pack_checksum[GIT_MAX_RAWSZ];
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
};
static struct bitmap_writer writer;
void bitmap_writer_show_progress(int show)
{
writer.show_progress = show;
}
/**
* Build the initial type index for the packfile
*/
void bitmap_writer_build_type_index(struct packing_data *to_pack,
struct pack_idx_entry **index,
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
uint32_t index_nr)
{
uint32_t i;
writer.commits = ewah_new();
writer.trees = ewah_new();
writer.blobs = ewah_new();
writer.tags = ewah_new();
ALLOC_ARRAY(to_pack->in_pack_pos, to_pack->nr_objects);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
for (i = 0; i < index_nr; ++i) {
struct object_entry *entry = (struct object_entry *)index[i];
enum object_type real_type;
oe_set_in_pack_pos(to_pack, entry, i);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
switch (oe_type(entry)) {
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
case OBJ_COMMIT:
case OBJ_TREE:
case OBJ_BLOB:
case OBJ_TAG:
real_type = oe_type(entry);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
break;
default:
real_type = oid_object_info(to_pack->repo,
&entry->idx.oid, NULL);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
break;
}
switch (real_type) {
case OBJ_COMMIT:
ewah_set(writer.commits, i);
break;
case OBJ_TREE:
ewah_set(writer.trees, i);
break;
case OBJ_BLOB:
ewah_set(writer.blobs, i);
break;
case OBJ_TAG:
ewah_set(writer.tags, i);
break;
default:
die("Missing type information for %s (%d/%d)",
oid_to_hex(&entry->idx.oid), real_type,
oe_type(entry));
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
}
}
/**
* Compute the actual bitmaps
*/
pack-bitmap-write: ignore BITMAP_FLAG_REUSE The on-disk bitmap format has a flag to mark a bitmap to be "reused". This is a rather curious feature, and works like this: - a run of pack-objects would decide to mark the last 80% of the bitmaps it generates with the reuse flag - the next time we generate bitmaps, we'd see those reuse flags from the last run, and mark those commits as special: - we'd be more likely to select those commits to get bitmaps in the new output - when generating the bitmap for a selected commit, we'd reuse the old bitmap as-is (rearranging the bits to match the new pack, of course) However, neither of these behaviors particularly makes sense. Just because a commit happened to be bitmapped last time does not make it a good candidate for having a bitmap this time. In particular, we may choose bitmaps based on how recent they are in history, or whether a ref tip points to them, and those things will change. We're better off re-considering fresh which commits are good candidates. Reusing the existing bitmap _is_ a reasonable thing to do to save computation. But only reusing exact bitmaps is a weak form of this. If we have an old bitmap for A and now want a new bitmap for its child, we should be able to compute that only by looking at trees and that are new to the child. But this code would consider only exact reuse (which is perhaps why it was eager to select those commits in the first place). Furthermore, the recent switch to the reverse-edge algorithm for generating bitmaps dropped this optimization entirely (and yet still performs better). So let's do a few cleanups: - drop the whole "reusing bitmaps" phase of generating bitmaps. It's not helping anything, and is mostly unused code (or worse, code that is using CPU but not doing anything useful) - drop the use of the on-disk reuse flag to select commits to bitmap - stop setting the on-disk reuse flag in bitmaps we generate (since nothing respects it anymore) We will keep a few innards of the reuse code, which will help us implement a more capable version of the "reuse" optimization: - simplify rebuild_existing_bitmaps() into a function that only builds the mapping of bits between the old and new orders, but doesn't actually convert any bitmaps - make rebuild_bitmap() public; we'll call it lazily to convert bitmaps as we traverse (using the mapping created above) Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:34 +00:00
static inline void push_bitmapped_commit(struct commit *commit)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
if (writer.selected_nr >= writer.selected_alloc) {
writer.selected_alloc = (writer.selected_alloc + 32) * 2;
REALLOC_ARRAY(writer.selected, writer.selected_alloc);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
writer.selected[writer.selected_nr].commit = commit;
pack-bitmap-write: ignore BITMAP_FLAG_REUSE The on-disk bitmap format has a flag to mark a bitmap to be "reused". This is a rather curious feature, and works like this: - a run of pack-objects would decide to mark the last 80% of the bitmaps it generates with the reuse flag - the next time we generate bitmaps, we'd see those reuse flags from the last run, and mark those commits as special: - we'd be more likely to select those commits to get bitmaps in the new output - when generating the bitmap for a selected commit, we'd reuse the old bitmap as-is (rearranging the bits to match the new pack, of course) However, neither of these behaviors particularly makes sense. Just because a commit happened to be bitmapped last time does not make it a good candidate for having a bitmap this time. In particular, we may choose bitmaps based on how recent they are in history, or whether a ref tip points to them, and those things will change. We're better off re-considering fresh which commits are good candidates. Reusing the existing bitmap _is_ a reasonable thing to do to save computation. But only reusing exact bitmaps is a weak form of this. If we have an old bitmap for A and now want a new bitmap for its child, we should be able to compute that only by looking at trees and that are new to the child. But this code would consider only exact reuse (which is perhaps why it was eager to select those commits in the first place). Furthermore, the recent switch to the reverse-edge algorithm for generating bitmaps dropped this optimization entirely (and yet still performs better). So let's do a few cleanups: - drop the whole "reusing bitmaps" phase of generating bitmaps. It's not helping anything, and is mostly unused code (or worse, code that is using CPU but not doing anything useful) - drop the use of the on-disk reuse flag to select commits to bitmap - stop setting the on-disk reuse flag in bitmaps we generate (since nothing respects it anymore) We will keep a few innards of the reuse code, which will help us implement a more capable version of the "reuse" optimization: - simplify rebuild_existing_bitmaps() into a function that only builds the mapping of bits between the old and new orders, but doesn't actually convert any bitmaps - make rebuild_bitmap() public; we'll call it lazily to convert bitmaps as we traverse (using the mapping created above) Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:34 +00:00
writer.selected[writer.selected_nr].bitmap = NULL;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
writer.selected[writer.selected_nr].flags = 0;
writer.selected_nr++;
}
static uint32_t find_object_pos(const struct object_id *oid)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
pack-objects: drop packlist index_pos optimization Once upon a time, the code to add an object to our packing list in pack-objects all lived in a single function. It computed the position within the hash table once, then used it to check if the object was already present, and if not, to add it. Later, in 2834bc27c1 (pack-objects: refactor the packing list, 2013-10-24), this was split into two functions: packlist_find() and packlist_alloc(). We ended up with an "index_pos" variable that gets passed through several functions to make it from one to the other. The resulting code is rather confusing to follow. The "index_pos" variable is sometimes undefined, if we don't yet have a hash table. This works out in practice because in that case packlist_alloc() won't use it at all, since it will have to create/grow the hash table. But it's hard to verify that, and it does cause gcc 9.2.1's -Wmaybe-uninitialized to complain when compiled with "-flto -O3" (rightfully, since we do pass the uninitialized value as a function parameter, even if nobody ends up using it). All of this is to save computing the hash index again when we're inserting into the hash table, which I found doesn't make a measurable difference in the program runtime (which is not surprising, since we're doing all kinds of other heavyweight things for each object). Let's just drop this index_pos variable entirely, simplifying the code (and pleasing the compiler). We might be better still refactoring this custom hash table to use one of our existing implementations (an oidmap, or a kh_oid_map). I stopped short of that here, but this would be the likely first step towards that anyway. Reported-by: Stephan Beyer <s-beyer@gmx.net> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-09-06 01:36:05 +00:00
struct object_entry *entry = packlist_find(writer.to_pack, oid);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
if (!entry) {
die("Failed to write bitmap index. Packfile doesn't have full closure "
"(object %s is missing)", oid_to_hex(oid));
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
return oe_in_pack_pos(writer.to_pack, entry);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
static void compute_xor_offsets(void)
{
static const int MAX_XOR_OFFSET_SEARCH = 10;
int i, next = 0;
while (next < writer.selected_nr) {
struct bitmapped_commit *stored = &writer.selected[next];
int best_offset = 0;
struct ewah_bitmap *best_bitmap = stored->bitmap;
struct ewah_bitmap *test_xor;
for (i = 1; i <= MAX_XOR_OFFSET_SEARCH; ++i) {
int curr = next - i;
if (curr < 0)
break;
test_xor = ewah_pool_new();
ewah_xor(writer.selected[curr].bitmap, stored->bitmap, test_xor);
if (test_xor->buffer_size < best_bitmap->buffer_size) {
if (best_bitmap != stored->bitmap)
ewah_pool_free(best_bitmap);
best_bitmap = test_xor;
best_offset = i;
} else {
ewah_pool_free(test_xor);
}
}
stored->xor_offset = best_offset;
stored->write_as = best_bitmap;
next++;
}
}
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
struct bb_commit {
struct commit_list *reverse_edges;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
struct bitmap *commit_mask;
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
struct bitmap *bitmap;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
unsigned selected:1,
maximal:1;
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
unsigned idx; /* within selected array */
};
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
define_commit_slab(bb_data, struct bb_commit);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
struct bitmap_builder {
struct bb_data data;
struct commit **commits;
size_t commits_nr, commits_alloc;
};
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
static void bitmap_builder_init(struct bitmap_builder *bb,
struct bitmap_writer *writer,
struct bitmap_index *old_bitmap)
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
{
struct rev_info revs;
struct commit *commit;
struct commit_list *reusable = NULL;
struct commit_list *r;
pack-bitmap-write: relax unique revwalk condition The previous commits improved the bitmap computation process for very long, linear histories with many refs by removing quadratic growth in how many objects were walked. The strategy of computing "intermediate commits" using bitmasks for which refs can reach those commits partitioned the poset of reachable objects so each part could be walked exactly once. This was effective for linear histories. However, there was a (significant) drawback: wide histories with many refs had an explosion of memory costs to compute the commit bitmasks during the exploration that discovers these intermediate commits. Since these wide histories are unlikely to repeat walking objects, the benefit of walking objects multiple times was not expensive before. But now, the commit walk *before computing bitmaps* is incredibly expensive. In an effort to discover a happy medium, this change reduces the walk for intermediate commits to only the first-parent history. This focuses the walk on how the histories converge, which still has significant reduction in repeat object walks. It is still possible to create quadratic behavior in this version, but it is probably less likely in realistic data shapes. Here is some data taken on a fresh clone of the kernel: | runtime (sec) | peak heap (GB) | | | | | from | with | from | with | | scratch | existing | scratch | existing | -----------+---------+----------+---------+----------- original | 64.044 | 83.241 | 2.088 | 2.194 | last patch | 45.049 | 37.624 | 2.267 | 2.334 | this patch | 88.478 | 53.218 | 2.157 | 2.224 | Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:05:26 +00:00
unsigned int i, num_maximal = 0;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
memset(bb, 0, sizeof(*bb));
init_bb_data(&bb->data);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
reset_revision_walk();
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
repo_init_revisions(writer->to_pack->repo, &revs, NULL);
revs.topo_order = 1;
pack-bitmap-write: relax unique revwalk condition The previous commits improved the bitmap computation process for very long, linear histories with many refs by removing quadratic growth in how many objects were walked. The strategy of computing "intermediate commits" using bitmasks for which refs can reach those commits partitioned the poset of reachable objects so each part could be walked exactly once. This was effective for linear histories. However, there was a (significant) drawback: wide histories with many refs had an explosion of memory costs to compute the commit bitmasks during the exploration that discovers these intermediate commits. Since these wide histories are unlikely to repeat walking objects, the benefit of walking objects multiple times was not expensive before. But now, the commit walk *before computing bitmaps* is incredibly expensive. In an effort to discover a happy medium, this change reduces the walk for intermediate commits to only the first-parent history. This focuses the walk on how the histories converge, which still has significant reduction in repeat object walks. It is still possible to create quadratic behavior in this version, but it is probably less likely in realistic data shapes. Here is some data taken on a fresh clone of the kernel: | runtime (sec) | peak heap (GB) | | | | | from | with | from | with | | scratch | existing | scratch | existing | -----------+---------+----------+---------+----------- original | 64.044 | 83.241 | 2.088 | 2.194 | last patch | 45.049 | 37.624 | 2.267 | 2.334 | this patch | 88.478 | 53.218 | 2.157 | 2.224 | Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:05:26 +00:00
revs.first_parent_only = 1;
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
for (i = 0; i < writer->selected_nr; i++) {
struct commit *c = writer->selected[i].commit;
struct bb_commit *ent = bb_data_at(&bb->data, c);
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
ent->selected = 1;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
ent->maximal = 1;
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
ent->idx = i;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
ent->commit_mask = bitmap_new();
bitmap_set(ent->commit_mask, i);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
add_pending_object(&revs, &c->object, "");
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
if (prepare_revision_walk(&revs))
die("revision walk setup failed");
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
while ((commit = get_revision(&revs))) {
pack-bitmap-write: relax unique revwalk condition The previous commits improved the bitmap computation process for very long, linear histories with many refs by removing quadratic growth in how many objects were walked. The strategy of computing "intermediate commits" using bitmasks for which refs can reach those commits partitioned the poset of reachable objects so each part could be walked exactly once. This was effective for linear histories. However, there was a (significant) drawback: wide histories with many refs had an explosion of memory costs to compute the commit bitmasks during the exploration that discovers these intermediate commits. Since these wide histories are unlikely to repeat walking objects, the benefit of walking objects multiple times was not expensive before. But now, the commit walk *before computing bitmaps* is incredibly expensive. In an effort to discover a happy medium, this change reduces the walk for intermediate commits to only the first-parent history. This focuses the walk on how the histories converge, which still has significant reduction in repeat object walks. It is still possible to create quadratic behavior in this version, but it is probably less likely in realistic data shapes. Here is some data taken on a fresh clone of the kernel: | runtime (sec) | peak heap (GB) | | | | | from | with | from | with | | scratch | existing | scratch | existing | -----------+---------+----------+---------+----------- original | 64.044 | 83.241 | 2.088 | 2.194 | last patch | 45.049 | 37.624 | 2.267 | 2.334 | this patch | 88.478 | 53.218 | 2.157 | 2.224 | Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:05:26 +00:00
struct commit_list *p = commit->parents;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
struct bb_commit *c_ent;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
parse_commit_or_die(commit);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
c_ent = bb_data_at(&bb->data, commit);
/*
* If there is no commit_mask, there is no reason to iterate
* over this commit; it is not selected (if it were, it would
* not have a blank commit mask) and all its children have
* existing bitmaps (see the comment starting with "This commit
* has an existing bitmap" below), so it does not contribute
* anything to the final bitmap file or its descendants.
*/
if (!c_ent->commit_mask)
continue;
if (old_bitmap && bitmap_for_commit(old_bitmap, commit)) {
/*
* This commit has an existing bitmap, so we can
* get its bits immediately without an object
* walk. That is, it is reusable as-is and there is no
* need to continue walking beyond it.
*
* Mark it as such and add it to bb->commits separately
* to avoid allocating a position in the commit mask.
*/
commit_list_insert(commit, &reusable);
goto next;
}
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
if (c_ent->maximal) {
pack-bitmap-write: relax unique revwalk condition The previous commits improved the bitmap computation process for very long, linear histories with many refs by removing quadratic growth in how many objects were walked. The strategy of computing "intermediate commits" using bitmasks for which refs can reach those commits partitioned the poset of reachable objects so each part could be walked exactly once. This was effective for linear histories. However, there was a (significant) drawback: wide histories with many refs had an explosion of memory costs to compute the commit bitmasks during the exploration that discovers these intermediate commits. Since these wide histories are unlikely to repeat walking objects, the benefit of walking objects multiple times was not expensive before. But now, the commit walk *before computing bitmaps* is incredibly expensive. In an effort to discover a happy medium, this change reduces the walk for intermediate commits to only the first-parent history. This focuses the walk on how the histories converge, which still has significant reduction in repeat object walks. It is still possible to create quadratic behavior in this version, but it is probably less likely in realistic data shapes. Here is some data taken on a fresh clone of the kernel: | runtime (sec) | peak heap (GB) | | | | | from | with | from | with | | scratch | existing | scratch | existing | -----------+---------+----------+---------+----------- original | 64.044 | 83.241 | 2.088 | 2.194 | last patch | 45.049 | 37.624 | 2.267 | 2.334 | this patch | 88.478 | 53.218 | 2.157 | 2.224 | Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:05:26 +00:00
num_maximal++;
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
ALLOC_GROW(bb->commits, bb->commits_nr + 1, bb->commits_alloc);
bb->commits[bb->commits_nr++] = commit;
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: relax unique revwalk condition The previous commits improved the bitmap computation process for very long, linear histories with many refs by removing quadratic growth in how many objects were walked. The strategy of computing "intermediate commits" using bitmasks for which refs can reach those commits partitioned the poset of reachable objects so each part could be walked exactly once. This was effective for linear histories. However, there was a (significant) drawback: wide histories with many refs had an explosion of memory costs to compute the commit bitmasks during the exploration that discovers these intermediate commits. Since these wide histories are unlikely to repeat walking objects, the benefit of walking objects multiple times was not expensive before. But now, the commit walk *before computing bitmaps* is incredibly expensive. In an effort to discover a happy medium, this change reduces the walk for intermediate commits to only the first-parent history. This focuses the walk on how the histories converge, which still has significant reduction in repeat object walks. It is still possible to create quadratic behavior in this version, but it is probably less likely in realistic data shapes. Here is some data taken on a fresh clone of the kernel: | runtime (sec) | peak heap (GB) | | | | | from | with | from | with | | scratch | existing | scratch | existing | -----------+---------+----------+---------+----------- original | 64.044 | 83.241 | 2.088 | 2.194 | last patch | 45.049 | 37.624 | 2.267 | 2.334 | this patch | 88.478 | 53.218 | 2.157 | 2.224 | Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:05:26 +00:00
if (p) {
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
struct bb_commit *p_ent = bb_data_at(&bb->data, p->item);
int c_not_p, p_not_c;
if (!p_ent->commit_mask) {
p_ent->commit_mask = bitmap_new();
c_not_p = 1;
p_not_c = 0;
} else {
c_not_p = bitmap_is_subset(c_ent->commit_mask, p_ent->commit_mask);
p_not_c = bitmap_is_subset(p_ent->commit_mask, c_ent->commit_mask);
}
if (!c_not_p)
continue;
bitmap_or(p_ent->commit_mask, c_ent->commit_mask);
if (p_not_c)
p_ent->maximal = 1;
else {
p_ent->maximal = 0;
free_commit_list(p_ent->reverse_edges);
p_ent->reverse_edges = NULL;
}
if (c_ent->maximal) {
commit_list_insert(commit, &p_ent->reverse_edges);
} else {
struct commit_list *cc = c_ent->reverse_edges;
for (; cc; cc = cc->next) {
if (!commit_list_contains(cc->item, p_ent->reverse_edges))
commit_list_insert(cc->item, &p_ent->reverse_edges);
}
}
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
}
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
next:
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
bitmap_free(c_ent->commit_mask);
c_ent->commit_mask = NULL;
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
}
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
for (r = reusable; r; r = r->next) {
ALLOC_GROW(bb->commits, bb->commits_nr + 1, bb->commits_alloc);
bb->commits[bb->commits_nr++] = r->item;
}
pack-bitmap-write: build fewer intermediate bitmaps The bitmap_writer_build() method calls bitmap_builder_init() to construct a list of commits reachable from the selected commits along with a "reverse graph". This reverse graph has edges pointing from a commit to other commits that can reach that commit. After computing a reachability bitmap for a commit, the values in that bitmap are then copied to the reachability bitmaps across the edges in the reverse graph. We can now relax the role of the reverse graph to greatly reduce the number of intermediate reachability bitmaps we compute during this reverse walk. The end result is that we walk objects the same number of times as before when constructing the reachability bitmaps, but we also spend much less time copying bits between bitmaps and have much lower memory pressure in the process. The core idea is to select a set of "important" commits based on interactions among the sets of commits reachable from each selected commit. The first technical concept is to create a new 'commit_mask' member in the bb_commit struct. Note that the selected commits are provided in an ordered array. The first thing to do is to mark the ith bit in the commit_mask for the ith selected commit. As we walk the commit-graph, we copy the bits in a commit's commit_mask to its parents. At the end of the walk, the ith bit in the commit_mask for a commit C stores a boolean representing "The ith selected commit can reach C." As we walk, we will discover non-selected commits that are important. We will get into this later, but those important commits must also receive bit positions, growing the width of the bitmasks as we walk. At the true end of the walk, the ith bit means "the ith _important_ commit can reach C." MAXIMAL COMMITS --------------- We use a new 'maximal' bit in the bb_commit struct to represent whether a commit is important or not. The term "maximal" comes from the partially-ordered set of commits in the commit-graph where C >= P if P is a parent of C, and then extending the relationship transitively. Instead of taking the maximal commits across the entire commit-graph, we instead focus on selecting each commit that is maximal among commits with the same bits on in their commit_mask. This definition is important, so let's consider an example. Suppose we have three selected commits A, B, and C. These are assigned bitmasks 100, 010, and 001 to start. Each of these can be marked as maximal immediately because they each will be the uniquely maximal commit that contains their own bit. Keep in mind that that these commits may have different bitmasks after the walk; for example, if B can reach C but A cannot, then the final bitmask for C is 011. Even in these cases, C would still be a maximal commit among all commits with the third bit on in their masks. Now define sets X, Y, and Z to be the sets of commits reachable from A, B, and C, respectively. The intersections of these sets correspond to different bitmasks: * 100: X - (Y union Z) * 010: Y - (X union Z) * 001: Z - (X union Y) * 110: (X intersect Y) - Z * 101: (X intersect Z) - Y * 011: (Y intersect Z) - X * 111: X intersect Y intersect Z This can be visualized with the following Hasse diagram: 100 010 001 | \ / \ / | | \/ \/ | | /\ /\ | | / \ / \ | 110 101 011 \___ | ___/ \ | / 111 Some of these bitmasks may not be represented, depending on the topology of the commit-graph. In fact, we are counting on it, since the number of possible bitmasks is exponential in the number of selected commits, but is also limited by the total number of commits. In practice, very few bitmasks are possible because most commits converge on a common "trunk" in the commit history. With this three-bit example, we wish to find commits that are maximal for each bitmask. How can we identify this as we are walking? As we walk, we visit a commit C. Since we are walking the commits in topo-order, we know that C is visited after all of its children are visited. Thus, when we get C from the revision walk we inspect the 'maximal' property of its bb_data and use that to determine if C is truly important. Its commit_mask is also nearly final. If C is not one of the originally-selected commits, then assign a bit position to C (by incrementing num_maximal) and set that bit on in commit_mask. See "MULTIPLE MAXIMAL COMMITS" below for more detail on this. Now that the commit C is known to be maximal or not, consider each parent P of C. Compute two new values: * c_not_p : true if and only if the commit_mask for C contains a bit that is not contained in the commit_mask for P. * p_not_c : true if and only if the commit_mask for P contains a bit that is not contained in the commit_mask for P. If c_not_p is false, then P already has all of the bits that C would provide to its commit_mask. In this case, move on to other parents as C has nothing to contribute to P's state that was not already provided by other children of P. We continue with the case that c_not_p is true. This means there are bits in C's commit_mask to copy to P's commit_mask, so use bitmap_or() to add those bits. If p_not_c is also true, then set the maximal bit for P to one. This means that if no other commit has P as a parent, then P is definitely maximal. This is because no child had the same bitmask. It is important to think about the maximal bit for P at this point as a temporary state: "P is maximal based on current information." In contrast, if p_not_c is false, then set the maximal bit for P to zero. Further, clear all reverse_edges for P since any edges that were previously assigned to P are no longer important. P will gain all reverse edges based on C. The final thing we need to do is to update the reverse edges for P. These reverse edges respresent "which closest maximal commits contributed bits to my commit_mask?" Since C contributed bits to P's commit_mask in this case, C must add to the reverse edges of P. If C is maximal, then C is a 'closest' maximal commit that contributed bits to P. Add C to P's reverse_edges list. Otherwise, C has a list of maximal commits that contributed bits to its bitmask (and this list is exactly one element). Add all of these items to P's reverse_edges list. Be careful to ignore duplicates here. After inspecting all parents P for a commit C, we can clear the commit_mask for C. This reduces the memory load to be limited to the "width" of the commit graph. Consider our ABC/XYZ example from earlier and let's inspect the state of the commits for an interesting bitmask, say 011. Suppose that D is the only maximal commit with this bitmask (in the first three bits). All other commits with bitmask 011 have D as the only entry in their reverse_edges list. D's reverse_edges list contains B and C. COMPUTING REACHABILITY BITMAPS ------------------------------ Now that we have our definition, let's zoom out and consider what happens with our new reverse graph when computing reachability bitmaps. We walk the reverse graph in reverse-topo-order, so we visit commits with largest commit_masks first. After we compute the reachability bitmap for a commit C, we push the bits in that bitmap to each commit D in the reverse edge list for C. Then, when we finally visit D we already have the bits for everything reachable from maximal commits that D can reach and we only need to walk the objects in the set-difference. In our ABC/XYZ example, when we finally walk for the commit A we only need to walk commits with bitmask equal to A's bitmask. If that bitmask is 100, then we are only walking commits in X - (Y union Z) because the bitmap already contains the bits for objects reachable from (X intersect Y) union (X intersect Z) (i.e. the bits from the reachability bitmaps for the maximal commits with bitmasks 110 and 101). The behavior is intended to walk each commit (and the trees that commit introduces) at most once while allocating and copying fewer reachability bitmaps. There is one caveat: what happens when there are multiple maximal commits with the same bitmask, with respect to the initial set of selected commits? MULTIPLE MAXIMAL COMMITS ------------------------ Earlier, we mentioned that when we discover a new maximal commit, we assign a new bit position to that commit and set that bit position to one for that commit. This is absolutely important for interesting commit-graphs such as git/git and torvalds/linux. The reason is due to the existence of "butterflies" in the commit-graph partial order. Here is an example of four commits forming a butterfly: I J |\ /| | \/ | | /\ | |/ \| M N \ / |/ Q Here, I and J both have parents M and N. In general, these do not need to be exact parent relationships, but reachability relationships. The most important part is that M and N cannot reach each other, so they are independent in the partial order. If I had commit_mask 10 and J had commit_mask 01, then M and N would both be assigned commit_mask 11 and be maximal commits with the bitmask 11. Then, what happens when M and N can both reach a commit Q? If Q is also assigned the bitmask 11, then it is not maximal but is reachable from both M and N. While this is not necessarily a deal-breaker for our abstract definition of finding maximal commits according to a given bitmask, we have a few issues that can come up in our larger picture of constructing reachability bitmaps. In particular, if we do not also consider Q to be a "maximal" commit, then we will walk commits reachable from Q twice: once when computing the reachability bitmap for M and another time when computing the reachability bitmap for N. This becomes much worse if the topology continues this pattern with multiple butterflies. The solution has already been mentioned: each of M and N are assigned their own bits to the bitmask and hence they become uniquely maximal for their bitmasks. Finally, Q also becomes maximal and thus we do not need to walk its commits multiple times. The final bitmasks for these commits are as follows: I:10 J:01 |\ /| | \ _____/ | | /\____ | |/ \ | M:111 N:1101 \ / Q:1111 Further, Q's reverse edge list is { M, N }, while M and N both have reverse edge list { I, J }. PERFORMANCE MEASUREMENTS ------------------------ Now that we've spent a LOT of time on the theory of this algorithm, let's show that this is actually worth all that effort. To test the performance, use GIT_TRACE2_PERF=1 when running 'git repack -abd' in a repository with no existing reachability bitmaps. This avoids any issues with keeping existing bitmaps to skew the numbers. Inspect the "building_bitmaps_total" region in the trace2 output to focus on the portion of work that is affected by this change. Here are the performance comparisons for a few repositories. The timings are for the following versions of Git: "multi" is the timing from before any reverse graph is constructed, where we might perform multiple traversals. "reverse" is for the previous change where the reverse graph has every reachable commit. Finally "maximal" is the version introduced here where the reverse graph only contains the maximal commits. Repository: git/git multi: 2.628 sec reverse: 2.344 sec maximal: 2.047 sec Repository: torvalds/linux multi: 64.7 sec reverse: 205.3 sec maximal: 44.7 sec So in all cases we've not only recovered any time lost to switching to the reverse-edge algorithm, but we come out ahead of "multi" in all cases. Likewise, peak heap has gone back to something reasonable: Repository: torvalds/linux multi: 2.087 GB reverse: 3.141 GB maximal: 2.288 GB While I do not have access to full fork networks on GitHub, Peff has run this algorithm on the chromium/chromium fork network and reported a change from 3 hours to ~233 seconds. That network is particularly beneficial for this approach because it has a long, linear history along with many tags. The "multi" approach was obviously quadratic and the new approach is linear. Helped-by: Jeff King <peff@peff.net> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:30 +00:00
trace2_data_intmax("pack-bitmap-write", the_repository,
"num_selected_commits", writer->selected_nr);
trace2_data_intmax("pack-bitmap-write", the_repository,
"num_maximal_commits", num_maximal);
free_commit_list(reusable);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
}
static void bitmap_builder_clear(struct bitmap_builder *bb)
{
clear_bb_data(&bb->data);
free(bb->commits);
bb->commits_nr = bb->commits_alloc = 0;
}
static void fill_bitmap_tree(struct bitmap *bitmap,
struct tree *tree)
{
uint32_t pos;
struct tree_desc desc;
struct name_entry entry;
/*
* If our bit is already set, then there is nothing to do. Both this
* tree and all of its children will be set.
*/
pos = find_object_pos(&tree->object.oid);
if (bitmap_get(bitmap, pos))
return;
bitmap_set(bitmap, pos);
if (parse_tree(tree) < 0)
die("unable to load tree object %s",
oid_to_hex(&tree->object.oid));
init_tree_desc(&desc, tree->buffer, tree->size);
while (tree_entry(&desc, &entry)) {
switch (object_type(entry.mode)) {
case OBJ_TREE:
fill_bitmap_tree(bitmap,
lookup_tree(the_repository, &entry.oid));
break;
case OBJ_BLOB:
bitmap_set(bitmap, find_object_pos(&entry.oid));
break;
default:
/* Gitlink, etc; not reachable */
break;
}
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
free_tree_buffer(tree);
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
static void fill_bitmap_commit(struct bb_commit *ent,
struct commit *commit,
struct prio_queue *queue,
struct prio_queue *tree_queue,
struct bitmap_index *old_bitmap,
const uint32_t *mapping)
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
{
if (!ent->bitmap)
ent->bitmap = bitmap_new();
prio_queue_put(queue, commit);
while (queue->nr) {
struct commit_list *p;
struct commit *c = prio_queue_get(queue);
if (old_bitmap && mapping) {
struct ewah_bitmap *old = bitmap_for_commit(old_bitmap, c);
/*
* If this commit has an old bitmap, then translate that
* bitmap and add its bits to this one. No need to walk
* parents or the tree for this commit.
*/
if (old && !rebuild_bitmap(mapping, old, ent->bitmap))
continue;
}
/*
* Mark ourselves and queue our tree. The commit
* walk ensures we cover all parents.
*/
bitmap_set(ent->bitmap, find_object_pos(&c->object.oid));
prio_queue_put(tree_queue, get_commit_tree(c));
for (p = c->parents; p; p = p->next) {
int pos = find_object_pos(&p->item->object.oid);
if (!bitmap_get(ent->bitmap, pos)) {
bitmap_set(ent->bitmap, pos);
prio_queue_put(queue, p->item);
}
}
}
while (tree_queue->nr)
fill_bitmap_tree(ent->bitmap, prio_queue_get(tree_queue));
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
static void store_selected(struct bb_commit *ent, struct commit *commit)
{
struct bitmapped_commit *stored = &writer.selected[ent->idx];
khiter_t hash_pos;
int hash_ret;
stored->bitmap = bitmap_to_ewah(ent->bitmap);
hash_pos = kh_put_oid_map(writer.bitmaps, commit->object.oid, &hash_ret);
if (hash_ret == 0)
die("Duplicate entry when writing index: %s",
oid_to_hex(&commit->object.oid));
kh_value(writer.bitmaps, hash_pos) = stored;
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
void bitmap_writer_build(struct packing_data *to_pack)
{
struct bitmap_builder bb;
size_t i;
int nr_stored = 0; /* for progress */
struct prio_queue queue = { compare_commits_by_gen_then_commit_date };
struct prio_queue tree_queue = { NULL };
struct bitmap_index *old_bitmap;
uint32_t *mapping;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
writer.bitmaps = kh_init_oid_map();
writer.to_pack = to_pack;
if (writer.show_progress)
writer.progress = start_progress("Building bitmaps", writer.selected_nr);
trace2_region_enter("pack-bitmap-write", "building_bitmaps_total",
the_repository);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
old_bitmap = prepare_bitmap_git(to_pack->repo);
if (old_bitmap)
mapping = create_bitmap_mapping(old_bitmap, to_pack);
else
mapping = NULL;
bitmap_builder_init(&bb, &writer, old_bitmap);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
for (i = bb.commits_nr; i > 0; i--) {
struct commit *commit = bb.commits[i-1];
struct bb_commit *ent = bb_data_at(&bb.data, commit);
struct commit *child;
pack-bitmap-write: pass ownership of intermediate bitmaps Our algorithm to generate reachability bitmaps walks through the commit graph from the bottom up, passing bitmap data from each commit to its descendants. For a linear stretch of history like: A -- B -- C our sequence of steps is: - compute the bitmap for A by walking its trees, etc - duplicate A's bitmap as a starting point for B; we can now free A's bitmap, since we only needed it as an intermediate result - OR in any extra objects that B can reach into its bitmap - duplicate B's bitmap as a starting point for C; likewise, free B's bitmap - OR in objects for C, and so on... Rather than duplicating bitmaps and immediately freeing the original, we can just pass ownership from commit to commit. Note that this doesn't always work: - the recipient may be a merge which already has an intermediate bitmap from its other ancestor. In that case we have to OR our result into it. Note that the first ancestor to reach the merge does get to pass ownership, though. - we may have multiple children; we can only pass ownership to one of them However, it happens often enough and copying bitmaps is expensive enough that this provides a noticeable speedup. On a clone of linux.git, this reduces the time to generate bitmaps from 205s to 70s. This is about the same amount of time it took to generate bitmaps using our old "many traversals" algorithm (the previous commit measures the identical scenario as taking 63s). It unfortunately provides only a very modest reduction in the peak memory usage, though. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:59 +00:00
int reused = 0;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
fill_bitmap_commit(ent, commit, &queue, &tree_queue,
old_bitmap, mapping);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
if (ent->selected) {
store_selected(ent, commit);
nr_stored++;
display_progress(writer.progress, nr_stored);
}
while ((child = pop_commit(&ent->reverse_edges))) {
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
struct bb_commit *child_ent =
bb_data_at(&bb.data, child);
if (child_ent->bitmap)
bitmap_or(child_ent->bitmap, ent->bitmap);
pack-bitmap-write: pass ownership of intermediate bitmaps Our algorithm to generate reachability bitmaps walks through the commit graph from the bottom up, passing bitmap data from each commit to its descendants. For a linear stretch of history like: A -- B -- C our sequence of steps is: - compute the bitmap for A by walking its trees, etc - duplicate A's bitmap as a starting point for B; we can now free A's bitmap, since we only needed it as an intermediate result - OR in any extra objects that B can reach into its bitmap - duplicate B's bitmap as a starting point for C; likewise, free B's bitmap - OR in objects for C, and so on... Rather than duplicating bitmaps and immediately freeing the original, we can just pass ownership from commit to commit. Note that this doesn't always work: - the recipient may be a merge which already has an intermediate bitmap from its other ancestor. In that case we have to OR our result into it. Note that the first ancestor to reach the merge does get to pass ownership, though. - we may have multiple children; we can only pass ownership to one of them However, it happens often enough and copying bitmaps is expensive enough that this provides a noticeable speedup. On a clone of linux.git, this reduces the time to generate bitmaps from 205s to 70s. This is about the same amount of time it took to generate bitmaps using our old "many traversals" algorithm (the previous commit measures the identical scenario as taking 63s). It unfortunately provides only a very modest reduction in the peak memory usage, though. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:59 +00:00
else if (reused)
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
child_ent->bitmap = bitmap_dup(ent->bitmap);
pack-bitmap-write: pass ownership of intermediate bitmaps Our algorithm to generate reachability bitmaps walks through the commit graph from the bottom up, passing bitmap data from each commit to its descendants. For a linear stretch of history like: A -- B -- C our sequence of steps is: - compute the bitmap for A by walking its trees, etc - duplicate A's bitmap as a starting point for B; we can now free A's bitmap, since we only needed it as an intermediate result - OR in any extra objects that B can reach into its bitmap - duplicate B's bitmap as a starting point for C; likewise, free B's bitmap - OR in objects for C, and so on... Rather than duplicating bitmaps and immediately freeing the original, we can just pass ownership from commit to commit. Note that this doesn't always work: - the recipient may be a merge which already has an intermediate bitmap from its other ancestor. In that case we have to OR our result into it. Note that the first ancestor to reach the merge does get to pass ownership, though. - we may have multiple children; we can only pass ownership to one of them However, it happens often enough and copying bitmaps is expensive enough that this provides a noticeable speedup. On a clone of linux.git, this reduces the time to generate bitmaps from 205s to 70s. This is about the same amount of time it took to generate bitmaps using our old "many traversals" algorithm (the previous commit measures the identical scenario as taking 63s). It unfortunately provides only a very modest reduction in the peak memory usage, though. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:59 +00:00
else {
child_ent->bitmap = ent->bitmap;
reused = 1;
}
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
}
pack-bitmap-write: pass ownership of intermediate bitmaps Our algorithm to generate reachability bitmaps walks through the commit graph from the bottom up, passing bitmap data from each commit to its descendants. For a linear stretch of history like: A -- B -- C our sequence of steps is: - compute the bitmap for A by walking its trees, etc - duplicate A's bitmap as a starting point for B; we can now free A's bitmap, since we only needed it as an intermediate result - OR in any extra objects that B can reach into its bitmap - duplicate B's bitmap as a starting point for C; likewise, free B's bitmap - OR in objects for C, and so on... Rather than duplicating bitmaps and immediately freeing the original, we can just pass ownership from commit to commit. Note that this doesn't always work: - the recipient may be a merge which already has an intermediate bitmap from its other ancestor. In that case we have to OR our result into it. Note that the first ancestor to reach the merge does get to pass ownership, though. - we may have multiple children; we can only pass ownership to one of them However, it happens often enough and copying bitmaps is expensive enough that this provides a noticeable speedup. On a clone of linux.git, this reduces the time to generate bitmaps from 205s to 70s. This is about the same amount of time it took to generate bitmaps using our old "many traversals" algorithm (the previous commit measures the identical scenario as taking 63s). It unfortunately provides only a very modest reduction in the peak memory usage, though. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:59 +00:00
if (!reused)
bitmap_free(ent->bitmap);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
ent->bitmap = NULL;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
clear_prio_queue(&queue);
clear_prio_queue(&tree_queue);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
bitmap_builder_clear(&bb);
free(mapping);
pack-bitmap-write: reimplement bitmap writing The bitmap generation code works by iterating over the set of commits for which we plan to write bitmaps, and then for each one performing a traditional traversal over the reachable commits and trees, filling in the bitmap. Between two traversals, we can often reuse the previous bitmap result as long as the first commit is an ancestor of the second. However, our worst case is that we may end up doing "n" complete complete traversals to the root in order to create "n" bitmaps. In a real-world case (the shared-storage repo consisting of all GitHub forks of chromium/chromium), we perform very poorly: generating bitmaps takes ~3 hours, whereas we can walk the whole object graph in ~3 minutes. This commit completely rewrites the algorithm, with the goal of accessing each object only once. It works roughly like this: - generate a list of commits in topo-order using a single traversal - invert the edges of the graph (so have parents point at their children) - make one pass in reverse topo-order, generating a bitmap for each commit and passing the result along to child nodes We generate correct results because each node we visit has already had all of its ancestors added to the bitmap. And we make only two linear passes over the commits. We also visit each tree usually only once. When filling in a bitmap, we don't bother to recurse into trees whose bit is already set in the bitmap (since we know we've already done so when setting their bit). That means that if commit A references tree T, none of its descendants will need to open T again. I say "usually", though, because it is possible for a given tree to be mentioned in unrelated parts of history (e.g., cherry-picking to a parallel branch). So we've accomplished our goal, and the resulting algorithm is pretty simple to understand. But there are some downsides, at least with this initial implementation: - we no longer reuse the results of any on-disk bitmaps when generating. So we'd expect to sometimes be slower than the original when bitmaps already exist. However, this is something we'll be able to add back in later. - we use much more memory. Instead of keeping one bitmap in memory at a time, we're passing them up through the graph. So our memory use should scale with the graph width (times the size of a bitmap). So how does it perform? For a clone of linux.git, generating bitmaps from scratch with the old algorithm took 63s. Using this algorithm it takes 205s. Which is much worse, but _might_ be acceptable if it behaved linearly as the size grew. It also increases peak heap usage by ~1G. That's not impossibly large, but not encouraging. On the complete fork-network of torvalds/linux, it increases the peak RAM usage by 40GB. Yikes. (I forgot to record the time it took, but the memory usage was too much to consider this reasonable anyway). On the complete fork-network of chromium/chromium, I ran out of memory before succeeding. Some back-of-the-envelope calculations indicate it would need 80+GB to complete. So at this stage, we've managed to make things much worse. But because of the way this new algorithm is structured, there are a lot of opportunities for optimization on top. We'll start implementing those in the follow-on patches. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:03:55 +00:00
trace2_region_leave("pack-bitmap-write", "building_bitmaps_total",
the_repository);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
stop_progress(&writer.progress);
compute_xor_offsets();
}
/**
* Select the commits that will be bitmapped
*/
static inline unsigned int next_commit_index(unsigned int idx)
{
static const unsigned int MIN_COMMITS = 100;
static const unsigned int MAX_COMMITS = 5000;
static const unsigned int MUST_REGION = 100;
static const unsigned int MIN_REGION = 20000;
unsigned int offset, next;
if (idx <= MUST_REGION)
return 0;
if (idx <= MIN_REGION) {
offset = idx - MUST_REGION;
return (offset < MIN_COMMITS) ? offset : MIN_COMMITS;
}
offset = idx - MIN_REGION;
next = (offset < MAX_COMMITS) ? offset : MAX_COMMITS;
return (next > MIN_COMMITS) ? next : MIN_COMMITS;
}
static int date_compare(const void *_a, const void *_b)
{
struct commit *a = *(struct commit **)_a;
struct commit *b = *(struct commit **)_b;
return (long)b->date - (long)a->date;
}
void bitmap_writer_select_commits(struct commit **indexed_commits,
unsigned int indexed_commits_nr,
int max_bitmaps)
{
unsigned int i = 0, j, next;
QSORT(indexed_commits, indexed_commits_nr, date_compare);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
if (writer.show_progress)
writer.progress = start_progress("Selecting bitmap commits", 0);
if (indexed_commits_nr < 100) {
for (i = 0; i < indexed_commits_nr; ++i)
pack-bitmap-write: ignore BITMAP_FLAG_REUSE The on-disk bitmap format has a flag to mark a bitmap to be "reused". This is a rather curious feature, and works like this: - a run of pack-objects would decide to mark the last 80% of the bitmaps it generates with the reuse flag - the next time we generate bitmaps, we'd see those reuse flags from the last run, and mark those commits as special: - we'd be more likely to select those commits to get bitmaps in the new output - when generating the bitmap for a selected commit, we'd reuse the old bitmap as-is (rearranging the bits to match the new pack, of course) However, neither of these behaviors particularly makes sense. Just because a commit happened to be bitmapped last time does not make it a good candidate for having a bitmap this time. In particular, we may choose bitmaps based on how recent they are in history, or whether a ref tip points to them, and those things will change. We're better off re-considering fresh which commits are good candidates. Reusing the existing bitmap _is_ a reasonable thing to do to save computation. But only reusing exact bitmaps is a weak form of this. If we have an old bitmap for A and now want a new bitmap for its child, we should be able to compute that only by looking at trees and that are new to the child. But this code would consider only exact reuse (which is perhaps why it was eager to select those commits in the first place). Furthermore, the recent switch to the reverse-edge algorithm for generating bitmaps dropped this optimization entirely (and yet still performs better). So let's do a few cleanups: - drop the whole "reusing bitmaps" phase of generating bitmaps. It's not helping anything, and is mostly unused code (or worse, code that is using CPU but not doing anything useful) - drop the use of the on-disk reuse flag to select commits to bitmap - stop setting the on-disk reuse flag in bitmaps we generate (since nothing respects it anymore) We will keep a few innards of the reuse code, which will help us implement a more capable version of the "reuse" optimization: - simplify rebuild_existing_bitmaps() into a function that only builds the mapping of bits between the old and new orders, but doesn't actually convert any bitmaps - make rebuild_bitmap() public; we'll call it lazily to convert bitmaps as we traverse (using the mapping created above) Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:34 +00:00
push_bitmapped_commit(indexed_commits[i]);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
return;
}
for (;;) {
struct commit *chosen = NULL;
next = next_commit_index(i);
if (i + next >= indexed_commits_nr)
break;
if (max_bitmaps > 0 && writer.selected_nr >= max_bitmaps) {
writer.selected_nr = max_bitmaps;
break;
}
if (next == 0) {
chosen = indexed_commits[i];
} else {
chosen = indexed_commits[i + next];
for (j = 0; j <= next; ++j) {
struct commit *cm = indexed_commits[i + j];
pack-bitmap-write: ignore BITMAP_FLAG_REUSE The on-disk bitmap format has a flag to mark a bitmap to be "reused". This is a rather curious feature, and works like this: - a run of pack-objects would decide to mark the last 80% of the bitmaps it generates with the reuse flag - the next time we generate bitmaps, we'd see those reuse flags from the last run, and mark those commits as special: - we'd be more likely to select those commits to get bitmaps in the new output - when generating the bitmap for a selected commit, we'd reuse the old bitmap as-is (rearranging the bits to match the new pack, of course) However, neither of these behaviors particularly makes sense. Just because a commit happened to be bitmapped last time does not make it a good candidate for having a bitmap this time. In particular, we may choose bitmaps based on how recent they are in history, or whether a ref tip points to them, and those things will change. We're better off re-considering fresh which commits are good candidates. Reusing the existing bitmap _is_ a reasonable thing to do to save computation. But only reusing exact bitmaps is a weak form of this. If we have an old bitmap for A and now want a new bitmap for its child, we should be able to compute that only by looking at trees and that are new to the child. But this code would consider only exact reuse (which is perhaps why it was eager to select those commits in the first place). Furthermore, the recent switch to the reverse-edge algorithm for generating bitmaps dropped this optimization entirely (and yet still performs better). So let's do a few cleanups: - drop the whole "reusing bitmaps" phase of generating bitmaps. It's not helping anything, and is mostly unused code (or worse, code that is using CPU but not doing anything useful) - drop the use of the on-disk reuse flag to select commits to bitmap - stop setting the on-disk reuse flag in bitmaps we generate (since nothing respects it anymore) We will keep a few innards of the reuse code, which will help us implement a more capable version of the "reuse" optimization: - simplify rebuild_existing_bitmaps() into a function that only builds the mapping of bits between the old and new orders, but doesn't actually convert any bitmaps - make rebuild_bitmap() public; we'll call it lazily to convert bitmaps as we traverse (using the mapping created above) Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:34 +00:00
if ((cm->object.flags & NEEDS_BITMAP) != 0) {
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
chosen = cm;
break;
}
if (cm->parents && cm->parents->next)
chosen = cm;
}
}
pack-bitmap-write: ignore BITMAP_FLAG_REUSE The on-disk bitmap format has a flag to mark a bitmap to be "reused". This is a rather curious feature, and works like this: - a run of pack-objects would decide to mark the last 80% of the bitmaps it generates with the reuse flag - the next time we generate bitmaps, we'd see those reuse flags from the last run, and mark those commits as special: - we'd be more likely to select those commits to get bitmaps in the new output - when generating the bitmap for a selected commit, we'd reuse the old bitmap as-is (rearranging the bits to match the new pack, of course) However, neither of these behaviors particularly makes sense. Just because a commit happened to be bitmapped last time does not make it a good candidate for having a bitmap this time. In particular, we may choose bitmaps based on how recent they are in history, or whether a ref tip points to them, and those things will change. We're better off re-considering fresh which commits are good candidates. Reusing the existing bitmap _is_ a reasonable thing to do to save computation. But only reusing exact bitmaps is a weak form of this. If we have an old bitmap for A and now want a new bitmap for its child, we should be able to compute that only by looking at trees and that are new to the child. But this code would consider only exact reuse (which is perhaps why it was eager to select those commits in the first place). Furthermore, the recent switch to the reverse-edge algorithm for generating bitmaps dropped this optimization entirely (and yet still performs better). So let's do a few cleanups: - drop the whole "reusing bitmaps" phase of generating bitmaps. It's not helping anything, and is mostly unused code (or worse, code that is using CPU but not doing anything useful) - drop the use of the on-disk reuse flag to select commits to bitmap - stop setting the on-disk reuse flag in bitmaps we generate (since nothing respects it anymore) We will keep a few innards of the reuse code, which will help us implement a more capable version of the "reuse" optimization: - simplify rebuild_existing_bitmaps() into a function that only builds the mapping of bits between the old and new orders, but doesn't actually convert any bitmaps - make rebuild_bitmap() public; we'll call it lazily to convert bitmaps as we traverse (using the mapping created above) Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-12-08 22:04:34 +00:00
push_bitmapped_commit(chosen);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
i += next + 1;
display_progress(writer.progress, i);
}
stop_progress(&writer.progress);
}
static int hashwrite_ewah_helper(void *f, const void *buf, size_t len)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
/* hashwrite will die on error */
hashwrite(f, buf, len);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
return len;
}
/**
* Write the bitmap index to disk
*/
static inline void dump_bitmap(struct hashfile *f, struct ewah_bitmap *bitmap)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
if (ewah_serialize_to(bitmap, hashwrite_ewah_helper, f) < 0)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
die("Failed to write bitmap index");
}
static const struct object_id *oid_access(size_t pos, const void *table)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
const struct pack_idx_entry * const *index = table;
return &index[pos]->oid;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}
static void write_selected_commits_v1(struct hashfile *f,
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
struct pack_idx_entry **index,
uint32_t index_nr)
{
int i;
for (i = 0; i < writer.selected_nr; ++i) {
struct bitmapped_commit *stored = &writer.selected[i];
int commit_pos =
oid_pos(&stored->commit->object.oid, index, index_nr, oid_access);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
if (commit_pos < 0)
BUG("trying to write commit not in index");
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
hashwrite_be32(f, commit_pos);
hashwrite_u8(f, stored->xor_offset);
hashwrite_u8(f, stored->flags);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
dump_bitmap(f, stored->write_as);
}
}
static void write_hash_cache(struct hashfile *f,
pack-bitmap: implement optional name_hash cache When we use pack bitmaps rather than walking the object graph, we end up with the list of objects to include in the packfile, but we do not know the path at which any tree or blob objects would be found. In a recently packed repository, this is fine. A fetch would use the paths only as a heuristic in the delta compression phase, and a fully packed repository should not need to do much delta compression. As time passes, though, we may acquire more objects on top of our large bitmapped pack. If clients fetch frequently, then they never even look at the bitmapped history, and all works as usual. However, a client who has not fetched since the last bitmap repack will have "have" tips in the bitmapped history, but "want" newer objects. The bitmaps themselves degrade gracefully in this circumstance. We manually walk the more recent bits of history, and then use bitmaps when we hit them. But we would also like to perform delta compression between the newer objects and the bitmapped objects (both to delta against what we know the user already has, but also between "new" and "old" objects that the user is fetching). The lack of pathnames makes our delta heuristics much less effective. This patch adds an optional cache of the 32-bit name_hash values to the end of the bitmap file. If present, a reader can use it to match bitmapped and non-bitmapped names during delta compression. Here are perf results for p5310: Test origin/master HEAD^ HEAD ------------------------------------------------------------------------------------------------- 5310.2: repack to disk 36.81(37.82+1.43) 47.70(48.74+1.41) +29.6% 47.75(48.70+1.51) +29.7% 5310.3: simulated clone 30.78(29.70+2.14) 1.08(0.97+0.10) -96.5% 1.07(0.94+0.12) -96.5% 5310.4: simulated fetch 3.16(6.10+0.08) 3.54(10.65+0.06) +12.0% 1.70(3.07+0.06) -46.2% 5310.6: partial bitmap 36.76(43.19+1.81) 6.71(11.25+0.76) -81.7% 4.08(6.26+0.46) -88.9% You can see that the time spent on an incremental fetch goes down, as our delta heuristics are able to do their work. And we save time on the partial bitmap clone for the same reason. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:45 +00:00
struct pack_idx_entry **index,
uint32_t index_nr)
{
uint32_t i;
for (i = 0; i < index_nr; ++i) {
struct object_entry *entry = (struct object_entry *)index[i];
hashwrite_be32(f, entry->hash);
pack-bitmap: implement optional name_hash cache When we use pack bitmaps rather than walking the object graph, we end up with the list of objects to include in the packfile, but we do not know the path at which any tree or blob objects would be found. In a recently packed repository, this is fine. A fetch would use the paths only as a heuristic in the delta compression phase, and a fully packed repository should not need to do much delta compression. As time passes, though, we may acquire more objects on top of our large bitmapped pack. If clients fetch frequently, then they never even look at the bitmapped history, and all works as usual. However, a client who has not fetched since the last bitmap repack will have "have" tips in the bitmapped history, but "want" newer objects. The bitmaps themselves degrade gracefully in this circumstance. We manually walk the more recent bits of history, and then use bitmaps when we hit them. But we would also like to perform delta compression between the newer objects and the bitmapped objects (both to delta against what we know the user already has, but also between "new" and "old" objects that the user is fetching). The lack of pathnames makes our delta heuristics much less effective. This patch adds an optional cache of the 32-bit name_hash values to the end of the bitmap file. If present, a reader can use it to match bitmapped and non-bitmapped names during delta compression. Here are perf results for p5310: Test origin/master HEAD^ HEAD ------------------------------------------------------------------------------------------------- 5310.2: repack to disk 36.81(37.82+1.43) 47.70(48.74+1.41) +29.6% 47.75(48.70+1.51) +29.7% 5310.3: simulated clone 30.78(29.70+2.14) 1.08(0.97+0.10) -96.5% 1.07(0.94+0.12) -96.5% 5310.4: simulated fetch 3.16(6.10+0.08) 3.54(10.65+0.06) +12.0% 1.70(3.07+0.06) -46.2% 5310.6: partial bitmap 36.76(43.19+1.81) 6.71(11.25+0.76) -81.7% 4.08(6.26+0.46) -88.9% You can see that the time spent on an incremental fetch goes down, as our delta heuristics are able to do their work. And we save time on the partial bitmap clone for the same reason. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:45 +00:00
}
}
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
void bitmap_writer_set_checksum(unsigned char *sha1)
{
hashcpy(writer.pack_checksum, sha1);
}
void bitmap_writer_finish(struct pack_idx_entry **index,
uint32_t index_nr,
pack-bitmap: implement optional name_hash cache When we use pack bitmaps rather than walking the object graph, we end up with the list of objects to include in the packfile, but we do not know the path at which any tree or blob objects would be found. In a recently packed repository, this is fine. A fetch would use the paths only as a heuristic in the delta compression phase, and a fully packed repository should not need to do much delta compression. As time passes, though, we may acquire more objects on top of our large bitmapped pack. If clients fetch frequently, then they never even look at the bitmapped history, and all works as usual. However, a client who has not fetched since the last bitmap repack will have "have" tips in the bitmapped history, but "want" newer objects. The bitmaps themselves degrade gracefully in this circumstance. We manually walk the more recent bits of history, and then use bitmaps when we hit them. But we would also like to perform delta compression between the newer objects and the bitmapped objects (both to delta against what we know the user already has, but also between "new" and "old" objects that the user is fetching). The lack of pathnames makes our delta heuristics much less effective. This patch adds an optional cache of the 32-bit name_hash values to the end of the bitmap file. If present, a reader can use it to match bitmapped and non-bitmapped names during delta compression. Here are perf results for p5310: Test origin/master HEAD^ HEAD ------------------------------------------------------------------------------------------------- 5310.2: repack to disk 36.81(37.82+1.43) 47.70(48.74+1.41) +29.6% 47.75(48.70+1.51) +29.7% 5310.3: simulated clone 30.78(29.70+2.14) 1.08(0.97+0.10) -96.5% 1.07(0.94+0.12) -96.5% 5310.4: simulated fetch 3.16(6.10+0.08) 3.54(10.65+0.06) +12.0% 1.70(3.07+0.06) -46.2% 5310.6: partial bitmap 36.76(43.19+1.81) 6.71(11.25+0.76) -81.7% 4.08(6.26+0.46) -88.9% You can see that the time spent on an incremental fetch goes down, as our delta heuristics are able to do their work. And we save time on the partial bitmap clone for the same reason. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:45 +00:00
const char *filename,
uint16_t options)
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
{
static uint16_t default_version = 1;
static uint16_t flags = BITMAP_OPT_FULL_DAG;
struct strbuf tmp_file = STRBUF_INIT;
struct hashfile *f;
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
struct bitmap_disk_header header;
int fd = odb_mkstemp(&tmp_file, "pack/tmp_bitmap_XXXXXX");
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
f = hashfd(fd, tmp_file.buf);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
memcpy(header.magic, BITMAP_IDX_SIGNATURE, sizeof(BITMAP_IDX_SIGNATURE));
header.version = htons(default_version);
pack-bitmap: implement optional name_hash cache When we use pack bitmaps rather than walking the object graph, we end up with the list of objects to include in the packfile, but we do not know the path at which any tree or blob objects would be found. In a recently packed repository, this is fine. A fetch would use the paths only as a heuristic in the delta compression phase, and a fully packed repository should not need to do much delta compression. As time passes, though, we may acquire more objects on top of our large bitmapped pack. If clients fetch frequently, then they never even look at the bitmapped history, and all works as usual. However, a client who has not fetched since the last bitmap repack will have "have" tips in the bitmapped history, but "want" newer objects. The bitmaps themselves degrade gracefully in this circumstance. We manually walk the more recent bits of history, and then use bitmaps when we hit them. But we would also like to perform delta compression between the newer objects and the bitmapped objects (both to delta against what we know the user already has, but also between "new" and "old" objects that the user is fetching). The lack of pathnames makes our delta heuristics much less effective. This patch adds an optional cache of the 32-bit name_hash values to the end of the bitmap file. If present, a reader can use it to match bitmapped and non-bitmapped names during delta compression. Here are perf results for p5310: Test origin/master HEAD^ HEAD ------------------------------------------------------------------------------------------------- 5310.2: repack to disk 36.81(37.82+1.43) 47.70(48.74+1.41) +29.6% 47.75(48.70+1.51) +29.7% 5310.3: simulated clone 30.78(29.70+2.14) 1.08(0.97+0.10) -96.5% 1.07(0.94+0.12) -96.5% 5310.4: simulated fetch 3.16(6.10+0.08) 3.54(10.65+0.06) +12.0% 1.70(3.07+0.06) -46.2% 5310.6: partial bitmap 36.76(43.19+1.81) 6.71(11.25+0.76) -81.7% 4.08(6.26+0.46) -88.9% You can see that the time spent on an incremental fetch goes down, as our delta heuristics are able to do their work. And we save time on the partial bitmap clone for the same reason. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:45 +00:00
header.options = htons(flags | options);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
header.entry_count = htonl(writer.selected_nr);
hashcpy(header.checksum, writer.pack_checksum);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
hashwrite(f, &header, sizeof(header) - GIT_MAX_RAWSZ + the_hash_algo->rawsz);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
dump_bitmap(f, writer.commits);
dump_bitmap(f, writer.trees);
dump_bitmap(f, writer.blobs);
dump_bitmap(f, writer.tags);
write_selected_commits_v1(f, index, index_nr);
pack-bitmap: implement optional name_hash cache When we use pack bitmaps rather than walking the object graph, we end up with the list of objects to include in the packfile, but we do not know the path at which any tree or blob objects would be found. In a recently packed repository, this is fine. A fetch would use the paths only as a heuristic in the delta compression phase, and a fully packed repository should not need to do much delta compression. As time passes, though, we may acquire more objects on top of our large bitmapped pack. If clients fetch frequently, then they never even look at the bitmapped history, and all works as usual. However, a client who has not fetched since the last bitmap repack will have "have" tips in the bitmapped history, but "want" newer objects. The bitmaps themselves degrade gracefully in this circumstance. We manually walk the more recent bits of history, and then use bitmaps when we hit them. But we would also like to perform delta compression between the newer objects and the bitmapped objects (both to delta against what we know the user already has, but also between "new" and "old" objects that the user is fetching). The lack of pathnames makes our delta heuristics much less effective. This patch adds an optional cache of the 32-bit name_hash values to the end of the bitmap file. If present, a reader can use it to match bitmapped and non-bitmapped names during delta compression. Here are perf results for p5310: Test origin/master HEAD^ HEAD ------------------------------------------------------------------------------------------------- 5310.2: repack to disk 36.81(37.82+1.43) 47.70(48.74+1.41) +29.6% 47.75(48.70+1.51) +29.7% 5310.3: simulated clone 30.78(29.70+2.14) 1.08(0.97+0.10) -96.5% 1.07(0.94+0.12) -96.5% 5310.4: simulated fetch 3.16(6.10+0.08) 3.54(10.65+0.06) +12.0% 1.70(3.07+0.06) -46.2% 5310.6: partial bitmap 36.76(43.19+1.81) 6.71(11.25+0.76) -81.7% 4.08(6.26+0.46) -88.9% You can see that the time spent on an incremental fetch goes down, as our delta heuristics are able to do their work. And we save time on the partial bitmap clone for the same reason. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:45 +00:00
if (options & BITMAP_OPT_HASH_CACHE)
write_hash_cache(f, index, index_nr);
finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
if (adjust_shared_perm(tmp_file.buf))
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
die_errno("unable to make temporary bitmap file readable");
if (rename(tmp_file.buf, filename))
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
die_errno("unable to rename temporary bitmap file to '%s'", filename);
strbuf_release(&tmp_file);
pack-objects: implement bitmap writing This commit extends more the functionality of `pack-objects` by allowing it to write out a `.bitmap` index next to any written packs, together with the `.idx` index that currently gets written. If bitmap writing is enabled for a given repository (either by calling `pack-objects` with the `--write-bitmap-index` flag or by having `pack.writebitmaps` set to `true` in the config) and pack-objects is writing a packfile that would normally be indexed (i.e. not piping to stdout), we will attempt to write the corresponding bitmap index for the packfile. Bitmap index writing happens after the packfile and its index has been successfully written to disk (`finish_tmp_packfile`). The process is performed in several steps: 1. `bitmap_writer_set_checksum`: this call stores the partial checksum for the packfile being written; the checksum will be written in the resulting bitmap index to verify its integrity 2. `bitmap_writer_build_type_index`: this call uses the array of `struct object_entry` that has just been sorted when writing out the actual packfile index to disk to generate 4 type-index bitmaps (one for each object type). These bitmaps have their nth bit set if the given object is of the bitmap's type. E.g. the nth bit of the Commits bitmap will be 1 if the nth object in the packfile index is a commit. This is a very cheap operation because the bitmap writing code has access to the metadata stored in the `struct object_entry` array, and hence the real type for each object in the packfile. 3. `bitmap_writer_reuse_bitmaps`: if there exists an existing bitmap index for one of the packfiles we're trying to repack, this call will efficiently rebuild the existing bitmaps so they can be reused on the new index. All the existing bitmaps will be stored in a `reuse` hash table, and the commit selection phase will prioritize these when selecting, as they can be written directly to the new index without having to perform a revision walk to fill the bitmap. This can greatly speed up the repack of a repository that already has bitmaps. 4. `bitmap_writer_select_commits`: if bitmap writing is enabled for a given `pack-objects` run, the sequence of commits generated during the Counting Objects phase will be stored in an array. We then use that array to build up the list of selected commits. Writing a bitmap in the index for each object in the repository would be cost-prohibitive, so we use a simple heuristic to pick the commits that will be indexed with bitmaps. The current heuristics are a simplified version of JGit's original implementation. We select a higher density of commits depending on their age: the 100 most recent commits are always selected, after that we pick 1 commit of each 100, and the gap increases as the commits grow older. On top of that, we make sure that every single branch that has not been merged (all the tips that would be required from a clone) gets their own bitmap, and when selecting commits between a gap, we tend to prioritize the commit with the most parents. Do note that there is no right/wrong way to perform commit selection; different selection algorithms will result in different commits being selected, but there's no such thing as "missing a commit". The bitmap walker algorithm implemented in `prepare_bitmap_walk` is able to adapt to missing bitmaps by performing manual walks that complete the bitmap: the ideal selection algorithm, however, would select the commits that are more likely to be used as roots for a walk in the future (e.g. the tips of each branch, and so on) to ensure a bitmap for them is always available. 5. `bitmap_writer_build`: this is the computationally expensive part of bitmap generation. Based on the list of commits that were selected in the previous step, we perform several incremental walks to generate the bitmap for each commit. The walks begin from the oldest commit, and are built up incrementally for each branch. E.g. consider this dag where A, B, C, D, E, F are the selected commits, and a, b, c, e are a chunk of simplified history that will not receive bitmaps. A---a---B--b--C--c--D \ E--e--F We start by building the bitmap for A, using A as the root for a revision walk and marking all the objects that are reachable until the walk is over. Once this bitmap is stored, we reuse the bitmap walker to perform the walk for B, assuming that once we reach A again, the walk will be terminated because A has already been SEEN on the previous walk. This process is repeated for C, and D, but when we try to generate the bitmaps for E, we can reuse neither the current walk nor the bitmap we have generated so far. What we do now is resetting both the walk and clearing the bitmap, and performing the walk from scratch using E as the origin. This new walk, however, does not need to be completed. Once we hit B, we can lookup the bitmap we have already stored for that commit and OR it with the existing bitmap we've composed so far, allowing us to limit the walk early. After all the bitmaps have been generated, another iteration through the list of commits is performed to find the best XOR offsets for compression before writing them to disk. Because of the incremental nature of these bitmaps, XORing one of them with its predecesor results in a minimal "bitmap delta" most of the time. We can write this delta to the on-disk bitmap index, and then re-compose the original bitmaps by XORing them again when loaded. This is a phase very similar to pack-object's `find_delta` (using bitmaps instead of objects, of course), except the heuristics have been greatly simplified: we only check the 10 bitmaps before any given one to find best compressing one. This gives good results in practice, because there is locality in the ordering of the objects (and therefore bitmaps) in the packfile. 6. `bitmap_writer_finish`: the last step in the process is serializing to disk all the bitmap data that has been generated in the two previous steps. The bitmap is written to a tmp file and then moved atomically to its final destination, using the same process as `pack-write.c:write_idx_file`. Signed-off-by: Vicent Marti <tanoku@gmail.com> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-12-21 14:00:16 +00:00
}