Borrow the bitmap index into packfiles from JGit to speed up
enumeration of objects involved in a commit range without having to
fully traverse the history.
* jk/pack-bitmap: (26 commits)
ewah: unconditionally ntohll ewah data
ewah: support platforms that require aligned reads
read-cache: use get_be32 instead of hand-rolled ntoh_l
block-sha1: factor out get_be and put_be wrappers
do not discard revindex when re-preparing packfiles
pack-bitmap: implement optional name_hash cache
t/perf: add tests for pack bitmaps
t: add basic bitmap functionality tests
count-objects: recognize .bitmap in garbage-checking
repack: consider bitmaps when performing repacks
repack: handle optional files created by pack-objects
repack: turn exts array into array-of-struct
repack: stop using magic number for ARRAY_SIZE(exts)
pack-objects: implement bitmap writing
rev-list: add bitmap mode to speed up object lists
pack-objects: use bitmaps when packing objects
pack-objects: split add_object_entry
pack-bitmap: add support for bitmap indexes
documentation: add documentation for the bitmap format
ewah: compressed bitmap implementation
...
Fix performance regression in v1.8.4.x and later.
* jk/mark-edges-uninteresting:
list-objects: only look at cmdline trees with edge_hint
t/perf: time rev-list with UNINTERESTING commits
We time a straight "rev-list --all" and its "--object"
counterpart, both going all the way to the root. However, we
do not time a partial history walk. This patch adds an
extreme case: a walk over a very small slice of history, but
with a very large set of UNINTERESTING tips. This is similar
to the connectivity check run by git on a small fetch, or
the walk done by any pre-receive hooks that want to check
incoming commits.
This test reveals a performance regression in git v1.8.4.2,
caused by fbd4a70 (list-objects: mark more commits as edges
in mark_edges_uninteresting, 2013-08-16):
Test fbd4a703^ fbd4a703
------------------------------------------------------------------------------------------
0001.1: rev-list --all 0.69(0.67+0.02) 0.69(0.68+0.01) +0.0%
0001.2: rev-list --all --objects 3.47(3.44+0.02) 3.48(3.44+0.03) +0.3%
0001.4: rev-list $commit --not --all 0.04(0.04+0.00) 0.04(0.04+0.00) +0.0%
0001.5: rev-list --objects $commit --not --all 0.04(0.03+0.00) 0.27(0.24+0.02) +575.0%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
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>
This adds a few basic perf tests for the pack bitmap code to
show off its improvements. The tests are:
1. How long does it take to do a repack (it gets slower
with bitmaps, since we have to do extra work)?
2. How long does it take to do a clone (it gets faster
with bitmaps)?
3. How does a small fetch perform when we've just
repacked?
4. How does a clone perform when we haven't repacked since
a week of pushes?
Here are results against linux.git:
Test origin/master this tree
-----------------------------------------------------------------------
5310.2: repack to disk 33.64(32.64+2.04) 67.67(66.75+1.84) +101.2%
5310.3: simulated clone 30.49(29.47+2.05) 1.20(1.10+0.10) -96.1%
5310.4: simulated fetch 3.49(6.79+0.06) 5.57(22.35+0.07) +59.6%
5310.6: partial bitmap 36.70(43.87+1.81) 8.18(21.92+0.73) -77.7%
You can see that we do take longer to repack, but we do way
better for further clones. A small fetch performs a bit
worse, as we spend way more time on delta compression (note
the heavy user CPU time, as we have 8 threads) due to the
lack of name hashes for the bitmapped objects.
The final test shows how the bitmaps degrade over time
between packs. There's still a significant speedup over the
non-bitmap case, but we don't do quite as well (we have to
spend time accessing the "new" objects the old fashioned
way, including delta compression).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git diff ../else/where/A ../else/where/B" when ../else/where is
clearly outside the repository, and "git diff --no-index A B", do
not have to look at the index at all, but we used to read the index
unconditionally.
* tg/diff-no-index-refactor:
diff: avoid some nesting
diff: add test for --no-index executed outside repo
diff: don't read index when --no-index is given
diff: move no-index detection to builtin/diff.c
git diff --no-index ... currently reads the index, during setup, when
calling gitmodules_config(). This results in worse performance when the
index is not actually needed. This patch avoids calling
gitmodules_config() when the --no-index option is given. The times for
executing "git diff --no-index" in the WebKit repository are improved as
follows:
Test HEAD~3 HEAD
------------------------------------------------------------------
4001.1: diff --no-index 0.24(0.15+0.09) 0.01(0.00+0.00) -95.8%
An additional improvement of this patch is that "git diff --no-index" no
longer breaks when the index file is corrupt, which makes it possible to
use it for investigating the broken repository.
To improve the possible usage as investigation tool for broken
repositories, setup_git_directory_gently() is also not called when the
--no-index option is given.
Also add a test to guard against future breakages, and a performance
test to show the improvements.
Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A #! line in these files is misleading, since these scriptlets are
meant to be sourced with '.' (using whatever shell sources them)
instead of run directly using the interpreter named on the #! line.
Removing the #! line shouldn't hurt syntax highlighting since
these files have filenames ending with '.sh'. For documentation,
add a brief description of how the files are meant to be used in
place of the shebang line.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
`echo -n` is non-portable. The POSIX specification says:
Conforming applications that wish to do prompting without <newline>
characters or that could possibly be expecting to echo a -n, should
use the printf utility derived from the Ninth Edition system.
Since all of the affected shell scripts use a POSIX shell shebang,
replace `echo -n` invocations with printf.
Signed-off-by: Lukas Fleischer <git@cryptocrack.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allows N instances of tests run in parallel, each running 1/N parts
of the test suite under Valgrind, to speed things up.
* tr/test-v-and-v-subtest-only:
perf-lib: fix start/stop of perf tests
test-lib: support running tests under valgrind in parallel
test-lib: allow prefixing a custom string before "ok N" etc.
test-lib: valgrind for only tests matching a pattern
test-lib: verbose mode for only tests matching a pattern
test-lib: self-test that --verbose works
test-lib: rearrange start/end of test_expect_* and test_skip
test-lib: refactor $GIT_SKIP_TESTS matching
test-lib: enable MALLOC_* for the actual tests
ae75342 test-lib: rearrange start/end of test_expect_* and test_skip
changed the way tests are started/stopped, but did not update the perf
tests. They were therefore giving the wrong output, because of the
wrong test count. Fix this by starting and stopping the tests
correctly.
Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
Acked-by: Thomas Rast <trast@inf.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 3.x tree has been out for a while now. The -2.6 repository name
survived the initial release [1], but kernel.org now only lists
'linux.git' (for aegl as well as torvalds) [2].
[1]: http://article.gmane.org/gmane.linux.kernel/1147422
On 2011-05-30 01:47:57 GMT, Linus Torvalds wrote:
> ... yes, that means that my git tree is still called
> "linux-2.6.git" on kernel.org.
[2]: http://git.kernel.org/cgit/
Signed-off-by: W. Trevor King <wking@tremily.us>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add the helper test-read-cache, which can be used to call read_cache and
discard_cache in a loop as well as a performance check based on it.
Signed-off-by: René Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* tr/line-log:
git-log(1): remove --full-line-diff description
line-log: fix documentation formatting
log -L: improve comments in process_all_files()
log -L: store the path instead of a diff_filespec
log -L: test merge of parallel modify/rename
t4211: pass -M to 'git log -M -L...' test
log -L: fix overlapping input ranges
log -L: check range set invariants when we look it up
Speed up log -L... -M
log -L: :pattern:file syntax to find by funcname
Implement line-history search (git log -L)
Export rewrite_parents() for 'log -L'
Refactor parse_loc
This is a rewrite of much of Bo's work, mainly in an effort to split
it into smaller, easier to understand routines.
The algorithm is built around the struct range_set, which encodes a
series of line ranges as intervals [a,b). This is used in two
contexts:
* A set of lines we are tracking (which will change as we dig through
history).
* To encode diffs, as pairs of ranges.
The main routine is range_set_map_across_diff(). It processes the
diff between a commit C and some parent P. It determines which diff
hunks are relevant to the ranges tracked in C, and computes the new
ranges for P.
The algorithm is then simply to process history in topological order
from newest to oldest, computing ranges and (partial) diffs. At
branch points, we need to merge the ranges we are watching. We will
find that many commits do not affect the chosen ranges, and mark them
TREESAME (in addition to those already filtered by pathspec limiting).
Another pass of history simplification then gets rid of such commits.
This is wired as an extra filtering pass in the log machinery. This
currently only reduces code duplication, but should allow for other
simplifications and options to be used.
Finally, we hook a diff printer into the output chain. Ideally we
would wire directly into the diff logic, to optionally use features
like word diff. However, that will require some major reworking of
the diff chain, so we completely replace the output with our own diff
for now.
As this was a GSoC project, and has quite some history by now, many
people have helped. In no particular order, thanks go to
Jakub Narebski <jnareb@gmail.com>
Jens Lehmann <Jens.Lehmann@web.de>
Jonathan Nieder <jrnieder@gmail.com>
Junio C Hamano <gitster@pobox.com>
Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Will Palmer <wmpalmer@gmail.com>
Apologies to everyone I forgot.
Signed-off-by: Bo Yang <struggleyb.nku@gmail.com>
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently the documentation of GIT_PERF_REPEAT_COUNT says the default is
five while "perf-lib.sh" uses a value of three as a default.
Update the documentation so that it is consistent with the code.
Signed-off-by: Antoine Pelisse <apelisse@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The malloc checks in tests are currently disabled. Actually evaluate
the variable for turning them off and enable them if it's unset.
Also use this opportunity to give it the more descriptive and
consistent name TEST_NO_MALLOC_CHECK.
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Run our test scripts with MALLOC_CHECK_ and MALLOC_PERTURB_, the
built-in memory access checking facility GNU libc has.
* ep/malloc-check-perturb:
MALLOC_CHECK: various clean-ups
Add MALLOC_CHECK_ and MALLOC_PERTURB_ libc env to the test suite for detecting heap corruption
The most important in this change is to avoid affecting anything
when test-lib is used from perf-lib. It also limits the effect of
the MALLOC_CHECK only to what is run inside the actual test, and
uses a fixed MALLOC_PERTURB_ in order to avoid hurting repeatability
of the tests.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King wrote:
The seq command is GNU-ism, and is missing at least in older BSD
releases and their derivatives, not to mention antique
commercial Unixes.
We already purged it in b3431bc (Don't use seq in tests, not
everyone has it, 2007-05-02), but a few new instances have crept
in. They went unnoticed because they are in scripts that are not
run by default.
Replace them with test_seq that is implemented with a Perl snippet
(proposed by Jeff). This is better than inlining this snippet
everywhere it's needed because it's easier to read and it's easier
to change the implementation (e.g. to C) if we ever decide to remove
Perl from the test suite.
Note that test_seq is not a complete replacement for seq(1). It
just has what we need now, in addition that it makes it possible for
us to do something like "test_seq a m" if we wanted to in the
future.
There are also many places that do `for i in 1 2 3 ...` but I'm not sure
if it's worth converting them to test_seq. That would introduce running
more processes of Perl.
Signed-off-by: Michał Kiedrowicz <michal.kiedrowicz@gmail.com>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Enables threading in index-pack to resolve base data in parallel.
By Nguyễn Thái Ngọc Duy (3) and Ramsay Jones (1)
* nd/threaded-index-pack:
index-pack: disable threading if NO_PREAD is defined
index-pack: support multithreaded delta resolving
index-pack: restructure pack processing into three main functions
compat/win32/pthread.h: Add an pthread_key_delete() implementation
This puts delta resolving on each base on a separate thread, one base
cache per thread. Per-thread data is grouped in struct thread_local.
When running with nr_threads == 1, no pthreads calls are made. The
system essentially runs in non-thread mode.
An experiment on a Xeon 24 core machine with git.git shows that
performance does not increase proportional to the number of cores. So
by default, we use maximum 3 cores. Some numbers with --threads from 1
to 16:
1..4
real 0m8.003s 0m5.307s 0m4.321s 0m3.830s
user 0m7.720s 0m8.009s 0m8.133s 0m8.305s
sys 0m0.224s 0m0.372s 0m0.360s 0m0.360s
5..8
real 0m3.727s 0m3.604s 0m3.332s 0m3.369s
user 0m9.361s 0m9.817s 0m9.525s 0m9.769s
sys 0m0.584s 0m0.624s 0m0.540s 0m0.560s
9..12
real 0m3.036s 0m3.139s 0m3.177s 0m2.961s
user 0m8.977s 0m10.205s 0m9.737s 0m10.073s
sys 0m0.596s 0m0.680s 0m0.684s 0m0.680s
13..16
real 0m2.985s 0m2.894s 0m2.975s 0m2.971s
user 0m9.825s 0m10.573s 0m10.833s 0m11.361s
sys 0m0.788s 0m0.732s 0m0.904s 0m1.016s
On an Intel dual core and linux-2.6.git
1..4
real 2m37.789s 2m7.963s 2m0.920s 1m58.213s
user 2m28.415s 2m52.325s 2m50.176s 2m41.187s
sys 0m7.808s 0m11.181s 0m11.224s 0m10.731s
Thanks Ramsay Jones for troubleshooting and support on MinGW platform.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'log -3000 (baseline)' test accidentally still used -1000 from an
earlier version.
Noticed-by: Lawrence Holding <Lawrence.Holding@cubic.com>
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The only bug right now is that $GIT_TEST_CMP is needed for test_cmp to
work.
However, we also export the three most important paths for tests:
TEST_DIRECTORY
TRASH_DIRECTORY
GIT_BUILD_DIR
Since they are available within test_expect_success, a future test
writer may expect them to also be defined in test_perf.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Loading it in the subshells still referred to $TEST_DIRECTORY/..,
which was only correct in preliminary versions of perf-lib.sh
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
8c912ee (teach --histogram to diff, 2011-07-12) claimed histogram diff
was faster than both Myers and patience.
We have since incorporated a performance testing framework, so add a
test that compares the various diff tasks performed in a real 'log -p'
workload. This does indeed show that histogram diff slightly beats
Myers, while patience is much slower than the others.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The only catch is that we don't really know what our repo contains, so
we have to ignore any possible "not found" status from git-grep.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This introduces a performance testing framework under t/perf/. It
tries to be as close to the test-lib.sh infrastructure as possible,
and thus should be easy to get used to for git developers.
The following points were considered for the implementation:
1. You usually want to compare arbitrary revisions/build trees against
each other. They may not have the performance test under
consideration, or even the perf-lib.sh infrastructure.
To cope with this, the 'run' script lets you specify arbitrary
build dirs and revisions. It even automatically builds the revisions
if it doesn't have them at hand yet.
2. Usually you would not want to run all tests. It would take too
long anyway. The 'run' script lets you specify which tests to run;
or you can also do it manually. There is a Makefile for
discoverability and 'make clean', but it is not meant for
real-world use.
3. Creating test repos from scratch in every test is extremely
time-consuming, and shipping or downloading such large/weird repos
is out of the question.
We leave this decision to the user. Two different sizes of test
repos can be configured, and the scripts just copy one or more of
those (using hardlinks for the object store). By default it tries
to use the build tree's git.git repository.
This is fairly fast and versatile. Using a copy instead of a clone
preserves many properties that the user may want to test for, such
as lots of loose objects, unpacked refs, etc.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
