Back in e37d0b8730 (builtin/index-pack.c: write reverse indexes,
2021-01-25), Git learned how to read and write a pack's reverse index
from a file instead of in-memory.
A pack's reverse index is a mapping from pack position (that is, the
order that objects appear together in a ".pack") to their position in
lexical order (that is, the order that objects are listed in an ".idx"
file).
Reverse indexes are consulted often during pack-objects, as well as
during auxiliary operations that require mapping between pack offsets,
pack order, and index index.
They are useful in GitHub's infrastructure, where we have seen a
dramatic increase in performance when writing ".rev" files[1]. In
particular:
- an ~80% reduction in the time it takes to serve fetches on a popular
repository, Homebrew/homebrew-core.
- a ~60% reduction in the peak memory usage to serve fetches on that
same repository.
- a collective savings of ~35% in CPU time across all pack-objects
invocations serving fetches across all repositories in a single
datacenter.
Reverse indexes are also beneficial to end-users as well as forges. For
example, the time it takes to generate a pack containing the objects for
the 10 most recent commits in linux.git (representing a typical push) is
significantly faster when on-disk reverse indexes are available:
$ { git rev-parse HEAD && printf '^' && git rev-parse HEAD~10 } >in
$ hyperfine -L v false,true 'git.compile -c pack.readReverseIndex={v} pack-objects --delta-base-offset --revs --stdout <in >/dev/null'
Benchmark 1: git.compile -c pack.readReverseIndex=false pack-objects --delta-base-offset --revs --stdout <in >/dev/null
Time (mean ± σ): 543.0 ms ± 20.3 ms [User: 616.2 ms, System: 58.8 ms]
Range (min … max): 521.0 ms … 577.9 ms 10 runs
Benchmark 2: git.compile -c pack.readReverseIndex=true pack-objects --delta-base-offset --revs --stdout <in >/dev/null
Time (mean ± σ): 245.0 ms ± 11.4 ms [User: 335.6 ms, System: 31.3 ms]
Range (min … max): 226.0 ms … 259.6 ms 13 runs
Summary
'git.compile -c pack.readReverseIndex=true pack-objects --delta-base-offset --revs --stdout <in >/dev/null' ran
2.22 ± 0.13 times faster than 'git.compile -c pack.readReverseIndex=false pack-objects --delta-base-offset --revs --stdout <in >/dev/null'
The same is true of writing a pack containing the objects for the 30
most-recent commits:
$ { git rev-parse HEAD && printf '^' && git rev-parse HEAD~30 } >in
$ hyperfine -L v false,true 'git.compile -c pack.readReverseIndex={v} pack-objects --delta-base-offset --revs --stdout <in >/dev/null'
Benchmark 1: git.compile -c pack.readReverseIndex=false pack-objects --delta-base-offset --revs --stdout <in >/dev/null
Time (mean ± σ): 866.5 ms ± 16.2 ms [User: 1414.5 ms, System: 97.0 ms]
Range (min … max): 839.3 ms … 886.9 ms 10 runs
Benchmark 2: git.compile -c pack.readReverseIndex=true pack-objects --delta-base-offset --revs --stdout <in >/dev/null
Time (mean ± σ): 581.6 ms ± 10.2 ms [User: 1181.7 ms, System: 62.6 ms]
Range (min … max): 567.5 ms … 599.3 ms 10 runs
Summary
'git.compile -c pack.readReverseIndex=true pack-objects --delta-base-offset --revs --stdout <in >/dev/null' ran
1.49 ± 0.04 times faster than 'git.compile -c pack.readReverseIndex=false pack-objects --delta-base-offset --revs --stdout <in >/dev/null'
...and savings on trivial operations like computing the on-disk size of
a single (packed) object are even more dramatic:
$ git rev-parse HEAD >in
$ hyperfine -L v false,true 'git.compile -c pack.readReverseIndex={v} cat-file --batch-check="%(objectsize:disk)" <in'
Benchmark 1: git.compile -c pack.readReverseIndex=false cat-file --batch-check="%(objectsize:disk)" <in
Time (mean ± σ): 305.8 ms ± 11.4 ms [User: 264.2 ms, System: 41.4 ms]
Range (min … max): 290.3 ms … 331.1 ms 10 runs
Benchmark 2: git.compile -c pack.readReverseIndex=true cat-file --batch-check="%(objectsize:disk)" <in
Time (mean ± σ): 4.0 ms ± 0.3 ms [User: 1.7 ms, System: 2.3 ms]
Range (min … max): 1.6 ms … 4.6 ms 1155 runs
Summary
'git.compile -c pack.readReverseIndex=true cat-file --batch-check="%(objectsize:disk)" <in' ran
76.96 ± 6.25 times faster than 'git.compile -c pack.readReverseIndex=false cat-file --batch-check="%(objectsize:disk)" <in'
In the more than two years since e37d0b8730 was merged, Git's
implementation of on-disk reverse indexes has been thoroughly tested,
both from users enabling `pack.writeReverseIndexes`, and from GitHub's
deployment of the feature. The latter has been running without incident
for more than two years.
This patch changes Git's behavior to write on-disk reverse indexes by
default when indexing a pack, which should make the above operations
faster for everybody's Git installation after a repack.
(The previous commit explains some potential drawbacks of using on-disk
reverse indexes in certain limited circumstances, that essentially boil
down to a trade-off between time to generate, and time to access. For
those limited cases, the `pack.readReverseIndex` escape hatch can be
used).
[1]: https://github.blog/2021-04-29-scaling-monorepo-maintenance/#reverse-indexes
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Acked-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git for-each-ref" learns '%(ahead-behind:<base>)' that computes the
distances from a single reference point in the history with bunch
of commits in bulk.
* ds/ahead-behind:
commit-reach: add tips_reachable_from_bases()
for-each-ref: add ahead-behind format atom
commit-reach: implement ahead_behind() logic
commit-graph: introduce `ensure_generations_valid()`
commit-graph: return generation from memory
commit-graph: simplify compute_generation_numbers()
commit-graph: refactor compute_topological_levels()
for-each-ref: explicitly test no matches
for-each-ref: add --stdin option
This argument was added in 7cae7627c4 (builtin/grep.c: integrate with
sparse index, 2022-09-22), but it was a carry-over from an earlier
version where the --sparse flag was added to the 'git grep' builtin.
This argument does not exist, so currently the
p2000-sparse-operations.sh performance test script fails when reaching
this step.
With this fix, the script works with these numbers for my copy of the
Git source code repository:
Test HEAD
------------------------------------------------------------
2000.30: git grep --cached ... (full-v3) 0.34(1.20+0.14)
2000.31: git grep --cached ... (full-v4) 0.31(1.15+0.13)
2000.32: git grep --cached ... (sparse-v3) 0.26(1.13+0.12)
2000.33: git grep --cached ... (sparse-v4) 0.27(1.13+0.12)
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Both 'git for-each-ref --merged=<X>' and 'git branch --merged=<X>' use
the ref-filter machinery to select references or branches (respectively)
that are reachable from a set of commits presented by one or more
--merged arguments. This happens within reach_filter(), which uses the
revision-walk machinery to walk history in a standard way.
However, the commit-reach.c file is full of custom searches that are
more efficient, especially for reachability queries that can terminate
early when reachability is discovered. Add a new
tips_reachable_from_bases() method to commit-reach.c and call it from
within reach_filter() in ref-filter.c. This affects both 'git branch'
and 'git for-each-ref' as tested in p1500-graph-walks.sh.
For the Linux kernel repository, we take an already-fast algorithm and
make it even faster:
Test HEAD~1 HEAD
-------------------------------------------------------------------
1500.5: contains: git for-each-ref --merged 0.13 0.02 -84.6%
1500.6: contains: git branch --merged 0.14 0.02 -85.7%
1500.7: contains: git tag --merged 0.15 0.03 -80.0%
(Note that we remove the iterative 'git rev-list' test from p1500
because it no longer makes sense as a comparison to 'git for-each-ref'
and would just waste time running it for these comparisons.)
The algorithm is implemented in commit-reach.c in the method
tips_reachable_from_base(). This method takes a string_list of tips and
assigns the 'util' for each item with the value 1 if the base commit can
reach those tips.
Like other reachability queries in commit-reach.c, the fastest way to
search for "can A reach B?" is to do a depth-first search up to the
generation number of B, preferring to explore first parents before later
parents. While we must walk all reachable commits up to that generation
number when the answer is "no", the depth-first search can answer "yes"
much faster than other approaches in most cases.
This search becomes trickier when there are multiple targets for the
depth-first search. The commits with lower generation number are more
likely to be within the history of the start commit, but we don't want
to waste time searching commits of low generation number if the commit
target with lowest generation number has already been found.
The trick here is to take the input commits and sort them by generation
number in ascending order. Track the index within this order as
min_generation_index. When we find a commit, if its index in the list is
equal to min_generation_index, then we can increase the generation
number boundary of our search to the next-lowest value in the list.
With this mechanism, the number of commits to search is minimized with
respect to the depth-first search heuristic. We will walk all commits up
to the minimum generation number of a commit that is _not_ reachable
from the start, but we will walk only the necessary portion of the
depth-first search for the reachable commits of lower generation.
Add extra tests for this behavior in t6600-test-reach.sh as the
interesting data shape of that repository can sometimes demonstrate
corner case bugs.
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous change implemented the ahead_behind() method, including an
algorithm to compute the ahead/behind values for a number of commit tips
relative to a number of commit bases. Now, integrate that algorithm as
part of 'git for-each-ref' hidden behind a new format atom,
ahead-behind. This naturally extends to 'git branch' and 'git tag'
builtins, as well.
This format allows specifying multiple bases, if so desired, and all
matching references are compared against all of those bases. For this
reason, failing to read a reference provided from these atoms results in
an error.
In order to translate the ahead_behind() method information to the
format output code in ref-filter.c, we must populate arrays of
ahead_behind_count structs. In struct ref_array, we store the full array
that will be passed to ahead_behind(). In struct ref_array_item, we
store an array of pointers that point to the relvant items within the
full array. In this way, we can pull all relevant ahead/behind values
directly when formatting output for a specific item. It also ensures the
lifetime of the ahead_behind_count structs matches the time that the
array is being used.
Add specific tests of the ahead/behind counts in t6600-test-reach.sh, as
it has an interesting repository shape. In particular, its merging
strategy and its use of different commit-graphs would demonstrate over-
counting if the ahead_behind() method did not already account for that
possibility.
Also add tests for the specific for-each-ref, branch, and tag builtins.
In the case of 'git tag', there are intersting cases that happen when
some of the selected tips are not commits. This requires careful logic
around commits_nr in the second loop of filter_ahead_behind(). Also, the
test in t7004 is carefully located to avoid being dependent on the GPG
prereq. It also avoids using the test_commit helper, as that will add
ticks to the time and disrupt the expected timestamps in later tag
tests.
Also add performance tests in a new p1300-graph-walks.sh script. This
will be useful for more uses in the future, but for now compare the
ahead-behind counting algorithm in 'git for-each-ref' to the naive
implementation by running 'git rev-list --count' processes for each
input.
For the Git source code repository, the improvement is already obvious:
Test this tree
---------------------------------------------------------------
1500.2: ahead-behind counts: git for-each-ref 0.07(0.07+0.00)
1500.3: ahead-behind counts: git branch 0.07(0.06+0.00)
1500.4: ahead-behind counts: git tag 0.07(0.06+0.00)
1500.5: ahead-behind counts: git rev-list 1.32(1.04+0.27)
But the standard performance benchmark is the Linux kernel repository,
which demosntrates a significant improvement:
Test this tree
---------------------------------------------------------------
1500.2: ahead-behind counts: git for-each-ref 0.27(0.24+0.02)
1500.3: ahead-behind counts: git branch 0.27(0.24+0.03)
1500.4: ahead-behind counts: git tag 0.28(0.27+0.01)
1500.5: ahead-behind counts: git rev-list 4.57(4.03+0.54)
The 'git rev-list' test exists in this change as a demonstration, but it
will be removed in the next change to avoid wasting time on this
comparison.
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When UTF is enabled for a PCRE match, the corresponding flags are
added to the pcre2_compile() call, but PCRE2_UCP wasn't included.
This prevents extending the meaning of the character classes to
include those new valid characters and therefore result in failed
matches for expressions that rely on that extention, for ex:
$ git grep -P '\bÆvar'
Add PCRE2_UCP so that \w will include Æ and therefore \b could
correctly match the beginning of that word.
This has an impact on performance that has been estimated to be
between 20% to 40% and that is shown through the added performance
test.
Signed-off-by: Carlo Marcelo Arenas Belón <carenas@gmail.com>
Acked-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When running 'read-tree' with a single tree and no prefix,
'prime_cache_tree()' is called after the tree is unpacked. In that
situation, skip a redundant call to 'cache_tree_update()' in
'unpack_trees()' by enabling the 'skip_cache_tree_update' unpack option.
Removing the redundant cache tree update provides a substantial performance
improvement to 'git read-tree <tree-ish>', as shown by a test added to
'p0006-read-tree-checkout.sh':
Test before after
----------------------------------------------------------------------
read-tree br_ballast_plus_1 3.94(1.80+1.57) 3.00(1.14+1.28) -23.9%
Note that the 'read-tree' in 't1022-read-tree-partial-clone.sh' is updated
to read two trees, rather than one. The test was first introduced in
d3da223f22 (cache-tree: prefetch in partial clone read-tree, 2021-07-23) to
exercise the 'cache_tree_update()' code path, as used in 'git merge'. Since
this patch drops the call to 'cache_tree_update()' in single-tree 'git
read-tree', change the test to use the two-tree variant so that
'cache_tree_update()' is called as intended.
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Enable the 'skip_cache_tree_update' option in the variants that call
'prime_cache_tree()' after 'unpack_trees()' (specifically, 'git reset
--mixed' and 'git reset --hard'). This avoids redundantly rebuilding the
cache tree in both 'cache_tree_update()' at the end of 'unpack_trees()' and
in 'prime_cache_tree()', resulting in a small (but consistent) performance
improvement. From the newly-added 'p7102-reset.sh' test:
Test before after
--------------------------------------------------------------------
7102.1: reset --hard (...) 2.11(0.40+1.54) 1.97(0.38+1.47) -6.6%
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Add a performance test comparing the execution times of 'prime_cache_tree()'
and 'cache_tree_update(_, WRITE_TREE_SILENT | WRITE_TREE_REPAIR)'. The goal
of comparing these two is to identify which is the faster method for
rebuilding an invalid cache tree, ultimately to remove one when both are
(reundantly) called in immediate succession.
Both methods are fast, so the new tests in 'p0090-cache-tree.sh' must call
each tested function multiple times to ensure the reported times (to 0.01s
resolution) convey the differences between them.
The tests compare the timing of a 'test-tool cache-tree' run as a no-op (to
capture a baseline for the overhead associated with running the tool),
'cache_tree_update()', and 'prime_cache_tree()' on four scenarios:
- A completely valid cache tree
- A cache tree with 2 invalid paths
- A cache tree with 50 invalid paths
- A completely empty cache tree
Example results:
Test this tree
-----------------------------------------------------------
0090.2: no-op, clean 1.27(0.48+0.52)
0090.3: prime_cache_tree, clean 2.02(0.83+0.85)
0090.4: cache_tree_update, clean 1.30(0.49+0.54)
0090.5: no-op, invalidate 2 1.29(0.48+0.54)
0090.6: prime_cache_tree, invalidate 2 1.98(0.81+0.83)
0090.7: cache_tree_update, invalidate 2 2.12(0.94+0.86)
0090.8: no-op, invalidate 50 1.32(0.50+0.55)
0090.9: prime_cache_tree, invalidate 50 2.10(0.86+0.89)
0090.10: cache_tree_update, invalidate 50 2.35(1.14+0.90)
0090.11: no-op, empty 1.33(0.50+0.54)
0090.12: prime_cache_tree, empty 2.04(0.84+0.87)
0090.13: cache_tree_update, empty 2.51(1.27+0.92)
These timings show that, while 'cache_tree_update()' is faster when the
cache tree is completely valid, it is equal to or slower than
'prime_cache_tree()' when there are any invalid paths. Since the redundant
calls are mostly in scenarios where the cache tree will be at least
partially invalid (e.g., 'git reset --hard'), 'prime_cache_tree()' will
likely perform better than 'cache_tree_update()' in typical cases.
Helped-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
"git grep" learned to expand the sparse-index more lazily and on
demand in a sparse checkout.
* sy/sparse-grep:
builtin/grep.c: integrate with sparse index
Turn on sparse index and remove ensure_full_index().
Before this patch, `git-grep` utilizes the ensure_full_index() method to
expand the index and search all the entries. Because this method
requires walking all the trees and constructing the index, it is the
slow part within the whole command.
To achieve better performance, this patch uses grep_tree() to search the
sparse directory entries and get rid of the ensure_full_index() method.
Why grep_tree() is a better choice over ensure_full_index()?
1) grep_tree() is as correct as ensure_full_index(). grep_tree() looks
into every sparse-directory entry (represented by a tree) recursively
when looping over the index, and the result of doing so matches the
result of expanding the index.
2) grep_tree() utilizes pathspecs to limit the scope of searching.
ensure_full_index() always expands the index, which means it will
always walk all the trees and blobs in the repo without caring if
the user only wants a subset of the content, i.e. using a pathspec.
On the other hand, grep_tree() will only search the contents that
match the pathspec, and thus possibly walking fewer trees.
3) grep_tree() does not construct and copy back a new index, while
ensure_full_index() does. This also saves some time.
----------------
Performance test
- Summary:
p2000 tests demonstrate a ~71% execution time reduction for
`git grep --cached bogus -- "f2/f1/f1/*"` using tree-walking logic.
However, notice that this result varies depending on the pathspec
given. See below "Command used for testing" for more details.
Test HEAD~ HEAD
-------------------------------------------------------
2000.78: git grep ... (full-v3) 0.35 0.39 (≈)
2000.79: git grep ... (full-v4) 0.36 0.30 (≈)
2000.80: git grep ... (sparse-v3) 0.88 0.23 (-73.8%)
2000.81: git grep ... (sparse-v4) 0.83 0.26 (-68.6%)
- Command used for testing:
git grep --cached bogus -- "f2/f1/f1/*"
The reason for specifying a pathspec is that, if we don't specify a
pathspec, then grep_tree() will walk all the trees and blobs to find the
pattern, and the time consumed doing so is not too different from using
the original ensure_full_index() method, which also spends most of the
time walking trees. However, when a pathspec is specified, this latest
logic will only walk the area of trees enclosed by the pathspec, and the
time consumed is reasonably a lot less.
Generally speaking, because the performance gain is acheived by walking
less trees, which are specified by the pathspec, the HEAD time v.s.
HEAD~ time in sparse-v[3|4], should be proportional to
"pathspec enclosed area" v.s. "all area", respectively. Namely, the
wider the <pathspec> is encompassing, the less the performance
difference between HEAD~ and HEAD, and vice versa.
That is, if we don't specify a pathspec, the performance difference [1]
is indistinguishable: both methods walk all the trees and take generally
same amount of time (even with the index construction time included for
ensure_full_index()).
[1] Performance test result without pathspec (hence walking all trees):
Command used:
git grep --cached bogus
Test HEAD~ HEAD
---------------------------------------------------
2000.78: git grep ... (full-v3) 6.17 5.19 (≈)
2000.79: git grep ... (full-v4) 6.19 5.46 (≈)
2000.80: git grep ... (sparse-v3) 6.57 6.44 (≈)
2000.81: git grep ... (sparse-v4) 6.65 6.28 (≈)
--------------------------
NEEDSWORK about submodules
There are a few NEEDSWORKs that belong to improvements beyond this
topic. See the NEEDSWORK in builtin/grep.c::grep_submodule() for
more context. The other two NEEDSWORKs in t1092 are also relative.
Suggested-by: Derrick Stolee <derrickstolee@github.com>
Helped-by: Derrick Stolee <derrickstolee@github.com>
Helped-by: Victoria Dye <vdye@github.com>
Helped-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Shaoxuan Yuan <shaoxuan.yuan02@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Despite POSIX states that:
> The old egrep and fgrep commands are likely to be supported for many
> years to come as implementation extensions, allowing historical
> applications to operate unmodified.
GNU grep 3.8 started to warn[1]:
> The egrep and fgrep commands, which have been deprecated since
> release 2.5.3 (2007), now warn that they are obsolescent and should
> be replaced by grep -E and grep -F.
Prepare for their removal in the future.
[1]: https://lists.gnu.org/archive/html/info-gnu/2022-09/msg00001.html
Signed-off-by: Đoàn Trần Công Danh <congdanhqx@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Hoist the remainder of "scalar" out of contrib/ to the main part of
the codebase.
* vd/scalar-to-main:
Documentation/technical: include Scalar technical doc
t/perf: add 'GIT_PERF_USE_SCALAR' run option
t/perf: add Scalar performance tests
scalar-clone: add test coverage
scalar: add to 'git help -a' command list
scalar: implement the `help` subcommand
git help: special-case `scalar`
scalar: include in standard Git build & installation
scalar: fix command documentation section header
The pack bitmap file gained a bitmap-lookup table to speed up
locating the necessary bitmap for a given commit.
* ac/bitmap-lookup-table:
pack-bitmap-write: drop unused pack_idx_entry parameters
bitmap-lookup-table: add performance tests for lookup table
pack-bitmap: prepare to read lookup table extension
pack-bitmap-write: learn pack.writeBitmapLookupTable and add tests
pack-bitmap-write.c: write lookup table extension
bitmap: move `get commit positions` code to `bitmap_writer_finish`
Documentation/technical: describe bitmap lookup table extension
Add a 'GIT_PERF_USE_SCALAR' environment variable (and corresponding perf
config 'useScalar') to register a repository created with any of:
* test_perf_fresh_repo
* test_perf_default_repo
* test_perf_large_repo
as a Scalar enlistment. This is intended to allow a developer to test the
impact of Scalar on already-defined performance scenarios.
Suggested-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Create 'p9210-scalar.sh' for testing Scalar performance and comparing
performance of Git operations in Scalar registrations and standard
repositories. Example results:
Test this tree
------------------------------------------------------------------------
9210.2: scalar clone 14.82(18.00+3.63)
9210.3: git clone 26.15(36.67+6.90)
9210.4: git status (scalar) 0.04(0.01+0.01)
9210.5: git status (non-scalar) 0.10(0.02+0.11)
9210.6: test_commit --append --no-tag A (scalar) 0.08(0.02+0.03)
9210.7: test_commit --append --no-tag A (non-scalar) 0.13(0.03+0.11)
Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix broken "&&-" chains and failures in early iterations of a loop.
* es/fix-chained-tests:
t5329: notice a failure within a loop
t: detect and signal failure within loop
t1092: fix buggy sparse "blame" test
t2407: fix broken &&-chains in compound statement
Add performance tests to verify the performance of lookup table.
`p5310-pack-bitmaps.sh` contain tests with and without lookup table.
`p5312-pack-bitmaps-revs.sh` contain same tests with and without
lookup table but with `pack.writeReverseIndex` enabled.
Lookup table makes Git run faster in most of the cases. Below is the
result of `t/perf/p5310-pack-bitmaps.sh`.`perf/p5326-multi-pack-bitmaps.sh`
gives similar result. The repository used in the test is linux kernel.
Test this tree
-----------------------------------------------------------------------
5310.4: enable lookup table: false 0.01(0.00+0.00)
5310.5: repack to disk 320.89(230.20+23.45)
5310.6: simulated clone 14.04(5.78+1.79)
5310.7: simulated fetch 1.95(3.05+0.20)
5310.8: pack to file (bitmap) 44.73(20.55+7.45)
5310.9: rev-list (commits) 0.78(0.46+0.10)
5310.10: rev-list (objects) 4.07(3.97+0.08)
5310.11: rev-list with tag negated via --not 0.06(0.02+0.03)
--all (objects)
5310.12: rev-list with negative tag (objects) 0.21(0.15+0.05)
5310.13: rev-list count with blob:none 0.24(0.17+0.06)
5310.14: rev-list count with blob:limit=1k 7.07(5.92+0.48)
5310.15: rev-list count with tree:0 0.25(0.17+0.07)
5310.16: simulated partial clone 5.67(3.28+0.64)
5310.18: clone (partial bitmap) 16.05(8.34+1.86)
5310.19: pack to file (partial bitmap) 59.76(27.22+7.43)
5310.20: rev-list with tree filter (partial bitmap) 0.90(0.18+0.16)
5310.24: enable lookup table: true 0.01(0.00+0.00)
5310.25: repack to disk 319.73(229.30+23.01)
5310.26: simulated clone 13.69(5.72+1.78)
5310.27: simulated fetch 1.84(3.02+0.16)
5310.28: pack to file (bitmap) 45.63(20.67+7.50)
5310.29: rev-list (commits) 0.56(0.39+0.8)
5310.30: rev-list (objects) 3.77(3.74+0.08)
5310.31: rev-list with tag negated via --not 0.05(0.02+0.03)
--all (objects)
5310.32: rev-list with negative tag (objects) 0.21(0.15+0.05)
5310.33: rev-list count with blob:none 0.23(0.17+0.05)
5310.34: rev-list count with blob:limit=1k 6.65(5.72+0.40)
5310.35: rev-list count with tree:0 0.23(0.16+0.06)
5310.36: simulated partial clone 5.57(3.26+0.59)
5310.38: clone (partial bitmap) 15.89(8.39+1.84)
5310.39: pack to file (partial bitmap) 58.32(27.55+7.47)
5310.40: rev-list with tree filter (partial bitmap) 0.73(0.18+0.15)
Test 4-15 are tested without using lookup table. Same tests are
repeated in 16-30 (using lookup table).
Mentored-by: Taylor Blau <me@ttaylorr.com>
Co-Mentored-by: Kaartic Sivaraam <kaartic.sivaraam@gmail.com>
Signed-off-by: Abhradeep Chakraborty <chakrabortyabhradeep79@gmail.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Failures within `for` and `while` loops can go unnoticed if not detected
and signaled manually since the loop itself does not abort when a
contained command fails, nor will a failure necessarily be detected when
the loop finishes since the loop returns the exit code of the last
command it ran on the final iteration, which may not be the command
which failed. Therefore, detect and signal failures manually within
loops using the idiom `|| return 1` (or `|| exit 1` within subshells).
Signed-off-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the 'p0006' test "read-tree br_base br_ballast", move the '-n' flag used
in 'git read-tree' ahead of its positional arguments.
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix multi-threaded 'p0004' test's use of the 'REPO_BIG_ENOUGH_FOR_MULTI'
prerequisite. Unlike normal 't/' tests, 't/perf/' tests need to have their
prerequisites declared with the '--prereq' flag.
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Enable the sparse index within the `git-rm` command.
The `p2000` tests demonstrate a ~92% execution time reduction for
'git rm' using a sparse index.
Test HEAD~1 HEAD
--------------------------------------------------------------------------
2000.74: git rm ... (full-v3) 0.41(0.37+0.05) 0.43(0.36+0.07) +4.9%
2000.75: git rm ... (full-v4) 0.38(0.34+0.05) 0.39(0.35+0.05) +2.6%
2000.76: git rm ... (sparse-v3) 0.57(0.56+0.01) 0.05(0.05+0.00) -91.2%
2000.77: git rm ... (sparse-v4) 0.57(0.55+0.02) 0.03(0.03+0.00) -94.7%
----
Also, normalize a behavioral difference of `git-rm` under sparse-index.
See related discussion [1].
`git-rm` a sparse-directory entry within a sparse-index enabled repo
behaves differently from a sparse directory within a sparse-checkout
enabled repo.
For example, in a sparse-index repo, where 'folder1' is a
sparse-directory entry, `git rm -r --sparse folder1` provides this:
rm 'folder1/'
Whereas in a sparse-checkout repo *without* sparse-index, doing so
provides this:
rm 'folder1/0/0/0'
rm 'folder1/0/1'
rm 'folder1/a'
Because `git rm` a sparse-directory entry does not need to expand the
index, therefore we should accept the current behavior, which is faster
than "expand the sparse-directory entry to match the sparse-checkout
situation".
Modify a previous test so such difference is not considered as an error.
[1] https://github.com/ffyuanda/git/pull/6#discussion_r934861398
Helped-by: Victoria Dye <vdye@github.com>
Helped-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Shaoxuan Yuan <shaoxuan.yuan02@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make our mergesort implementation type-safe.
* rs/mergesort:
mergesort: remove llist_mergesort()
packfile: use DEFINE_LIST_SORT
fetch-pack: use DEFINE_LIST_SORT
commit: use DEFINE_LIST_SORT
blame: use DEFINE_LIST_SORT
test-mergesort: use DEFINE_LIST_SORT
test-mergesort: use DEFINE_LIST_SORT_DEBUG
mergesort: add macros for typed sort of linked lists
mergesort: tighten merge loop
mergesort: unify ranks loops
Build a typed sort function for the mergesort performance test tool
using DEFINE_LIST_SORT instead of calling llist_mergesort(). This gets
rid of the next pointer accessor functions and improves the performance
at the cost of a slightly higher object text size.
Before:
0071.12: llist_mergesort() unsorted 0.24(0.22+0.01)
0071.14: llist_mergesort() sorted 0.12(0.10+0.01)
0071.16: llist_mergesort() reversed 0.12(0.10+0.01)
__TEXT __DATA __OBJC others dec hex
6407 276 0 24701 31384 7a98 t/helper/test-mergesort.o
With this patch:
0071.12: DEFINE_LIST_SORT unsorted 0.22(0.21+0.01)
0071.14: DEFINE_LIST_SORT sorted 0.11(0.10+0.01)
0071.16: DEFINE_LIST_SORT reversed 0.11(0.10+0.01)
__TEXT __DATA __OBJC others dec hex
6615 276 0 25832 32723 7fd3 t/helper/test-mergesort.o
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 5dccd9155f (t/perf: add iteration setup mechanism to perf-lib,
2022-04-04) modified the parameter parsing of test_wrapper() such that
the test title was no longer in $1, and is instead in $test_title_.
We correctly pass the new variable to the code which outputs the title
to the log, but missed the spot in test_wrapper() where the title is
written to the ".descr" file which is used to produce the final output
table. As a result, all of the titles are missing from that table (or
worse, using whatever was left in $1):
$ ./p0000-perf-lib-sanity.sh
[...]
Test this tree
------------------------------
0000.1: 0.01(0.01+0.00)
0000.2: 0.01(0.00+0.01)
0000.4: 0.00(0.00+0.00)
0000.5: true 0.00(0.00+0.00)
0000.7: 0.00(0.00+0.00)
0000.8: 0.00(0.00+0.00)
After this patch, we get the pre-5dccd9155f output:
Test this tree
--------------------------------------------------------------------------
0000.1: test_perf_default_repo works 0.00(0.00+0.00)
0000.2: test_checkout_worktree works 0.01(0.00+0.01)
0000.4: export a weird var 0.00(0.00+0.00)
0000.5: éḿíẗ ńöń-ÁŚĆÍÍ ćḧáŕáćẗéŕś 0.00(0.00+0.00)
0000.7: important variables available in subshells 0.00(0.00+0.00)
0000.8: test-lib-functions correctly loaded in subshells 0.00(0.00+0.00)
Signed-off-by: Jeff King <peff@peff.net>
Acked-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
More fsmonitor--daemon.
* jh/builtin-fsmonitor-part3: (30 commits)
t7527: improve implicit shutdown testing in fsmonitor--daemon
fsmonitor--daemon: allow --super-prefix argument
t7527: test Unicode NFC/NFD handling on MacOS
t/lib-unicode-nfc-nfd: helper prereqs for testing unicode nfc/nfd
t/helper/hexdump: add helper to print hexdump of stdin
fsmonitor: on macOS also emit NFC spelling for NFD pathname
t7527: test FSMonitor on case insensitive+preserving file system
fsmonitor: never set CE_FSMONITOR_VALID on submodules
t/perf/p7527: add perf test for builtin FSMonitor
t7527: FSMonitor tests for directory moves
fsmonitor: optimize processing of directory events
fsm-listen-darwin: shutdown daemon if worktree root is moved/renamed
fsm-health-win32: force shutdown daemon if worktree root moves
fsm-health-win32: add polling framework to monitor daemon health
fsmonitor--daemon: stub in health thread
fsmonitor--daemon: rename listener thread related variables
fsmonitor--daemon: prepare for adding health thread
fsmonitor--daemon: cd out of worktree root
fsm-listen-darwin: ignore FSEvents caused by xattr changes on macOS
unpack-trees: initialize fsmonitor_has_run_once in o->result
...
"sparse-checkout" learns to work well with the sparse-index
feature.
* ds/sparse-sparse-checkout:
sparse-checkout: integrate with sparse index
p2000: add test for 'git sparse-checkout [add|set]'
sparse-index: complete partial expansion
sparse-index: partially expand directories
sparse-checkout: --no-sparse-index needs a full index
cache-tree: implement cache_tree_find_path()
sparse-index: introduce partially-sparse indexes
sparse-index: create expand_index()
t1092: stress test 'git sparse-checkout set'
t1092: refactor 'sparse-index contents' test
Introduce a filesystem-dependent mechanism to optimize the way the
bits for many loose object files are ensured to hit the disk
platter.
* ns/batch-fsync:
core.fsyncmethod: performance tests for batch mode
t/perf: add iteration setup mechanism to perf-lib
core.fsyncmethod: tests for batch mode
test-lib-functions: add parsing helpers for ls-files and ls-tree
core.fsync: use batch mode and sync loose objects by default on Windows
unpack-objects: use the bulk-checkin infrastructure
update-index: use the bulk-checkin infrastructure
builtin/add: add ODB transaction around add_files_to_cache
cache-tree: use ODB transaction around writing a tree
core.fsyncmethod: batched disk flushes for loose-objects
bulk-checkin: rebrand plug/unplug APIs as 'odb transactions'
bulk-checkin: rename 'state' variable and separate 'plugged' boolean
The sparse-checkout builtin is almost completely integrated with the
sparse index, allowing the sparse-checkout boundary to be modified
without expanding a sparse index to a full one. Add a test to
p2000-sparse-operations.sh that adds a directory to the sparse-checkout
definition, then removes it. Using both operations is important to
ensure that the operation is doing the same work in each repetition as
well as leaving the test repo in a good state for later tests.
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add tests verifying expected 'git stash' behavior in
't1092-sparse-checkout-compatibility'. These cases establish the expected
behavior of 'git stash' in a sparse-checkout and verify consistency both
with and without a sparse index. Although no sparse index compatibility has
been integrated into 'git stash' yet, the tests are all 'expect_success' -
we don't want the cone-mode sparse-checkout behavior to change depending on
whether it is using a sparse index or not. Therefore, we expect these tests
to continue passing once sparse index is integrated with 'git stash'.
Additionally, add performance test cases for 'git stash' both with and
without untracked files. Note that, unlike the other tests in
'p2000-sparse-operations.sh', the tests added for 'stash' are combination
operations. This is done to ensure the stash/unstash is not blocked by the
modification of '$SPARSE_CONE/a' performed as part of 'test_perf_on_all'.
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Tests that affect the repo in stateful ways are easier to write if we
can run setup steps outside of the measured portion of perf iteration.
This change adds a "--setup 'setup-script'" parameter to test_perf. To
make invocations easier to understand, I also moved the prerequisites to
a new --prereq parameter.
The setup facility will be used in the upcoming perf tests for batch
mode, but it already helps in some existing tests, like t5302 and t7820.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add basic performance tests for git commands that can add data to the
object database. We cover:
* git add
* git stash
* git update-index (via git stash)
* git unpack-objects
* git commit --all
We cover all currently available fsync methods as well.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Built-in fsmonitor (part 2).
* jh/builtin-fsmonitor-part2: (30 commits)
t7527: test status with untracked-cache and fsmonitor--daemon
fsmonitor: force update index after large responses
fsmonitor--daemon: use a cookie file to sync with file system
fsmonitor--daemon: periodically truncate list of modified files
t/perf/p7519: add fsmonitor--daemon test cases
t/perf/p7519: speed up test on Windows
t/perf/p7519: fix coding style
t/helper/test-chmtime: skip directories on Windows
t/perf: avoid copying builtin fsmonitor files into test repo
t7527: create test for fsmonitor--daemon
t/helper/fsmonitor-client: create IPC client to talk to FSMonitor Daemon
help: include fsmonitor--daemon feature flag in version info
fsmonitor--daemon: implement handle_client callback
compat/fsmonitor/fsm-listen-darwin: implement FSEvent listener on MacOS
compat/fsmonitor/fsm-listen-darwin: add MacOS header files for FSEvent
compat/fsmonitor/fsm-listen-win32: implement FSMonitor backend on Windows
fsmonitor--daemon: create token-based changed path cache
fsmonitor--daemon: define token-ids
fsmonitor--daemon: add pathname classification
fsmonitor--daemon: implement 'start' command
...
Repeat all of the fsmonitor perf tests using `git fsmonitor--daemon` and
the "Simple IPC" interface.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change p7519 to use `test_seq` and `xargs` rather than a `for` loop
to touch thousands of files. This takes minutes off of test runs
on Windows because of process creation overhead.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Do not copy any of the various fsmonitor--daemon files from the .git
directory of the (GIT_PREF_REPO or GIT_PERF_LARGE_REPO) source repo
into the test's trash directory.
When perf tests start, they copy the contents of the source repo into
the test's trash directory. If fsmonitor is running in the source repo,
there may be control files, such as the IPC socket and/or fsmonitor
cookie files. These should not be copied into the test repo.
Unix domain sockets cannot be copied in the manner used by the test
setup, so if present, the test setup fails.
Cookie files are harmless, but we should avoid them.
The builtin fsmonitor keeps all such control files/sockets in
.git/fsmonitor--daemon*, so it is simple to exclude them.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Optimize away strbuf_expand() call with a hardcoded formatting logic
specific for the default format in the --batch and --batch-check
options of "git cat-file".
* jc/cat-file-batch-default-format-optim:
cat-file: skip expanding default format
"git read-tree" has been made to be aware of the sparse-index
feature.
* vd/sparse-read-tree:
read-tree: make three-way merge sparse-aware
read-tree: make two-way merge sparse-aware
read-tree: narrow scope of index expansion for '--prefix'
read-tree: integrate with sparse index
read-tree: expand sparse checkout test coverage
read-tree: explicitly disallow prefixes with a leading '/'
status: fix nested sparse directory diff in sparse index
sparse-index: prevent repo root from becoming sparse
When format is passed into --batch, --batch-check, --batch-command,
the format gets expanded. When nothing is passed in, the default format
is set and the expand_format() gets called.
We can save on these cycles by hardcoding how to print the
information when nothing is passed as the format, or when the default
format is passed. There is no need for the fully expanded format with
the default. Since batch_object_write() happens on every object provided
in batch mode, we get a nice performance improvement.
git rev-list --all > /tmp/all-obj.txt
git cat-file --batch-check </tmp/all-obj.txt
with HEAD^:
Time (mean ± σ): 57.6 ms ± 1.7 ms [User: 51.5 ms, System: 6.2 ms]
Range (min … max): 54.6 ms … 64.7 ms 50 runs
with HEAD:
Time (mean ± σ): 49.8 ms ± 1.7 ms [User: 42.6 ms, System: 7.3 ms]
Range (min … max): 46.9 ms … 55.9 ms 56 runs
If nothing is provided as a format argument, or if the default format is
passed, skip expanding of the format and print the object info with a
default format.
See https://lore.kernel.org/git/87eecf8ork.fsf@evledraar.gmail.com/
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: John Cai <johncai86@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We have various behavior that's shared across our Makefiles, or that
really should be (e.g. via defined templates). Let's create a
top-level "shared.mak" to house those sorts of things, and start by
adding the ".DELETE_ON_ERROR" flag to it.
See my own 7b76d6bf22 (Makefile: add and use the ".DELETE_ON_ERROR"
flag, 2021-06-29) and db10fc6c09 (doc: simplify Makefile using
.DELETE_ON_ERROR, 2021-05-21) for the addition and use of the
".DELETE_ON_ERROR" flag.
I.e. this changes the behavior of existing rules in the altered
Makefiles (except "Makefile" & "Documentation/Makefile"). I'm
confident that this is safe having read the relevant rules in those
Makfiles, and as the GNU make manual notes that it isn't the default
behavior is out of an abundance of backwards compatibility
caution. From edition 0.75 of its manual, covering GNU make 4.3:
[Enabling '.DELETE_ON_ERROR' is] almost always what you want
'make' to do, but it is not historical practice; so for
compatibility, you must explicitly request it.
This doesn't introduce a bug by e.g. having this
".DELETE_ON_ERROR" flag only apply to this new shared.mak, Makefiles
have no such scoping semantics.
It does increase the danger that any Makefile without an explicit "The
default target of this Makefile is..." snippet to define the default
target as "all" could have its default rule changed if our new
shared.mak ever defines a "real" rule. In subsequent commits we'll be
careful not to do that, and such breakage would be obvious e.g. in the
case of "make -C t".
We might want to make that less fragile still (e.g. by using
".DEFAULT_GOAL" as noted in the preceding commit), but for now let's
simply include "shared.mak" without adding that boilerplate to all the
Makefiles that don't have it already. Most of those are already
exposed to that potential caveat e.g. due to including "config.mak*".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add tests focused on how 'git read-tree' behaves in sparse checkouts. Extra
emphasis is placed on interactions with files outside the sparse cone, e.g.
merges with out-of-cone conflicts.
Helped-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Victoria Dye <vdye@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git update-index", "git checkout-index", and "git clean" are
taught to work better with the sparse checkout feature.
* vd/sparse-clean-etc:
update-index: reduce scope of index expansion in do_reupdate
update-index: integrate with sparse index
update-index: add tests for sparse-checkout compatibility
checkout-index: integrate with sparse index
checkout-index: add --ignore-skip-worktree-bits option
checkout-index: expand sparse checkout compatibility tests
clean: integrate with sparse index
reset: reorder wildcard pathspec conditions
reset: fix validation in sparse index test
Introduce tests for a variety of `git update-index` use cases, including
performance scenarios. Tests are intended to exercise `update-index` with
options that change the commands interaction with the index (e.g.,
`--again`) and with files/directories inside and outside a sparse checkout
cone.
Of note is that these tests clearly establish the behavior of `git
update-index --add` with untracked, outside-of-cone files. Unlike `git add`,
which fails with an error when provided with such files, `update-index`
succeeds in adding them to the index. Additionally, the `skip-worktree` flag
is *not* automatically added to the new entry. Although this is pre-existing
behavior, there are a couple of reasons to avoid changing it in favor of
consistency with e.g. `git add`:
* `update-index` is low-level command for modifying the index; while it can
perform operations similar to those of `add`, it traditionally has fewer
"guardrails" preventing a user from doing something they may not want to
do (in this case, adding an outside-of-cone, non-`skip-worktree` file to
the index)
* `update-index` typically only exits with an error code if it is incapable
of performing an operation (e.g., if an internal function call fails);
adding a new file outside the sparse checkout definition is still a valid
operation, albeit an inadvisable one
* `update-index` does not implicitly set flags (e.g., `skip-worktree`) when
creating new index entries with `--add`; if flags need to be updated,
options like `--[no-]skip-worktree` allow a user to intentionally set them
All this to say that, while there are valid reasons to consider changing the
treatment of outside-of-cone files in `update-index`, there are also
sufficient reasons for leaving it as-is.
Co-authored-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Victoria Dye <vdye@github.com>
Reviewed-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add tests to cover `checkout-index`, with a focus on cases interesting in a
sparse checkout (e.g., files specified outside sparse checkout definition).
New tests are intended to serve as a baseline for existing and/or expected
behavior and performance when integrating `checkout-index` with the sparse
index. Note that the test 'checkout-index --all' is marked as
'test_expect_failure', indicating that `update-index --all` will be modified
in a subsequent patch to behave as the test expects.
Signed-off-by: Victoria Dye <vdye@github.com>
Reviewed-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Perf tests were run with end-user's shell, but it has been
corrected to use the shell specified by $TEST_SHELL_PATH.
* ja/perf-use-specified-shell:
t/perf: do not run tests in user's $SHELL