The performance of adding numerous new packfiles has been improved
by introducing a fast path for known-new packfiles to skip an
unnecessary traversal in packfile_list_append(), avoiding a
quadratic complexity regression on load.
* js/packfile-fast-append:
packfile: fix perf regression with many packs
Since 589127caa7 (packfile: move list of packs into the packfile
store, 2025-10-30), there is a performance regression when many
packfiles need to be loaded: `packfile_store_add_pack()` now calls
`packfile_list_remove_internal()` to detect whether the packfile was
_already_ in the list, and if so, move it to the end of the list. This
function linearly scans the existing list before every insertion. Newly
loading N packs therefore has complexity O(N²).
In one reported use case (https://github.com/microsoft/git/issues/970),
N equals 37,815 and caused a slow-down of a simple `git rev-parse
--short HEAD` (which is regularly executed as part of `GIT_PS1`) from
0.4s to 4.5s.
Let's fix this by establishing a fast path for known-new packfiles.
The keen reader will note that there is currently only a single,
"known-new" caller of the `packfile_list_append()` function, and wonder
why not simply remove this check whether the packfile already exists in
the list? Originally, when above-mentioned commit introduced that logic,
there was a second caller in `prepare_midx()`, which would have required
that check, but that caller was removed in 6aff1f25a0 (packfile:
always add packfiles to MRU when adding a pack, 2025-10-30). Still, the
function is declared in a header file, and to avoid any problems with
in-flight or downstream callers, it is safer to extend the signature to
be explicit whether or not to skip that check.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The 'git stash push' command has been optimized to avoid unnecessary
sparse index expansion when pathspecs are wholly inside the
sparse-checkout cone. Also, a potential out-of-bounds read in the
sparse-index expansion check helper pathspec_needs_expanded_index()
has been fixed by consistently using the parsed, prefixed path.
* tn/stash-avoid-sparse-index-expansion:
stash: avoid sparse-index expansion for in-cone paths
pathspec: use match for sparse-index expansion checks
`git stash push -- <pathspec>` expands a sparse index before checking
whether the pathspec matches any tracked paths. This is unnecessary
when the pathspec is wholly inside the sparse-checkout cone and makes
a path-limited stash proportional to the size of the full index.
Use `pathspec_needs_expanded_index()` to expand only when a pathspec
can match part of a sparse-directory entry, as `git rm` and `git
reset` already do. Keep the full-index behavior for pathspecs that
need it.
Add compatibility coverage for literal, prefixed, wildcard, file,
multiple, staged, and missing pathspecs. Add the corresponding
path-limited stash case to p2000.
On a cone-mode repository with 349,525 tracked paths and 49 sparse
index entries, the best of three runs changed from 18.87s to 0.06s.
Trace2 reported four index expansions before this change and none
after it.
Signed-off-by: Ted Nyman <tnyman@openai.com>
Reviewed-by: Taylor Blau <ttaylorr@openai.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
'git branch --contains' and 'git for-each-ref --contains' have been
optimized to use the memoized commit traversal previously used only by
'git tag --contains', significantly speeding up connectivity checks
across many candidate refs with shared history.
* td/ref-filter-memoize-contains:
commit-reach: die on contains walk errors
ref-filter: memoize --contains with generations
commit-reach: reject cycles in contains walk
The performance of ref updates and reads using the 'reftable' backend
in the presence of many deletion tombstone records has been optimized
by removing the tombstone suppression flag from the merged iterator
and instead skipping tombstones at higher-level call sites where
iteration bounds are known.
* kk/reftable-tombstone-quadratic-fix:
reftable: fix quadratic behavior in the presence of tombstones
t/perf: add perf test for ref tombstone scenarios
The cache-scanning loop in 'next_cache_entry()' has been optimized
to avoid rescanning already-unpacked index entries, preventing a
quadratic performance slow-down when diffing the working tree
against a commit with a pathspec matching early index entries.
* hf/unpack-trees-quadratic-scan:
unpack-trees: avoid quadratic index scan in next_cache_entry()
Add performance tests for update-ref when many tombstones are present
in a reftable.
The first test exercises the scenario where all refs are deleted
(creating tombstones) and then re-created with the same names, which
currently exhibits quadratic behavior.
The second test uses a separate repository with an asymmetric variant
where refs are deleted and then new, differently-named refs are
created. When the tombstones sort after the new refs, every create
scans all tombstones, making this case even worse than re-creating
the same refs.
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Kristofer Karlsson <krka@spotify.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Diffing the working tree against a commit with a pathspec can take
time quadratic in the size of the index when the pathspec matches a
subtree whose entries are the first entries of the index. Fix it by
having next_cache_entry() record how far it scanned in cache_bottom,
so repeated calls no longer rescan the growing prefix of
already-unpacked entries. On a Chromium checkout (~500k index
entries),
git diff HEAD -- .agents/OWNERS
took about 8 minutes before this change and 0.07 seconds after it.
The same diff without the commit, without the pathspec, or with
--cached was already instant.
Add p0009-diff-pathspec.sh, which builds a 10,000-entry index whose
first path lives in a subtree (100,000 entries under --long-tests),
to guard against the regression. Comparing v2.55.0 with this change
using GIT_TEST_LONG=t:
Test v2.55.0 HEAD
------------------------------------------------------------------------
0009.2: diff pathspec subtree 7.16(7.12+0.01) 0.02(0.01+0.00) -99.7%
Signed-off-by: Henrique Ferreiro <hferreiro@igalia.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The pack-objects command has been updated to support reachability
bitmaps and delta-islands concurrently with the `--path-walk` option,
allowing faster packaging by falling back to path-walk when bitmaps
cannot fully satisfy the request.
* tb/pack-path-walk-bitmap-delta-islands:
pack-objects: support `--delta-islands` with `--path-walk`
pack-objects: extract `record_tree_depth()` helper
pack-objects: support reachability bitmaps with `--path-walk`
t/perf: drop p5311's lookup-table permutation
Commands that list branches and tags (like git branch and git tag)
have been optimized to pass the namespace prefix when initializing
their ref iterator, avoiding a loose-ref scaling regression in
repositories with many unrelated loose references.
* td/ref-filter-restore-prefix-iteration:
ref-filter: restore prefix-scoped iteration
When 'pack-objects' is invoked with '--path-walk', it prevents us from
using reachability bitmaps.
This behavior dates back to 70664d2865 (pack-objects: add --path-walk
option, 2025-05-16), which included a comment in the relevant portion of
the command-line arguments handling that read as follows:
/*
* We must disable the bitmaps because we are removing
* the --objects / --objects-edge[-aggressive] options.
*/
In fb2c309b7d3 (pack-objects: pass --objects with --path-walk,
2026-05-02), path-walk learned to pass '--objects' again, but still
kept bitmap traversal disabled. That leaves two useful cases
unsupported:
* A path-walk repack that writes bitmaps does not give the bitmap
selector any commits, because path-walk reveals commits through
`add_objects_by_path()` rather than through `show_commit()`, where
`index_commit_for_bitmap()` is normally called.
* An invocation like "git pack-objects --use-bitmap-index --path-walk"
never tries an existing bitmap, even when one is available and could
answer the request.
Fortunately for us, neither restriction is required.
* On the writing side: teach the path-walk object callback to call
`index_commit_for_bitmap()` for commits that it adds to the pack.
That gives the bitmap selector the commit candidates it would have
seen from the regular traversal.
* For bitmap reading, keep passing '--objects' to the internal rev_list
machinery, but stop clearing `use_bitmap_index`. If an existing
bitmap can answer the request, use it; otherwise fall back to
path-walk's own enumeration.
As a result, we can see significantly reduced pack generation times from
p5311 (with our `GIT_PERF_REPO` set to a recent clone of the fluentui
repository) before this commit:
Test HEAD^ HEAD
----------------------------------------------------------------------------------------
5311.40: server (1 days, --path-walk) 1.43(1.39+0.04) 0.01(0.01+0.00) -99.3%
5311.41: size (1 days, --path-walk) 139.6K 139.7K +0.0%
5311.42: client (1 days, --path-walk) 0.02(0.02+0.00) 0.02(0.02+0.00) +0.0%
5311.44: server (2 days, --path-walk) 1.43(1.39+0.04) 0.01(0.00+0.00) -99.3%
5311.45: size (2 days, --path-walk) 139.6K 139.7K +0.0%
5311.46: client (2 days, --path-walk) 0.02(0.02+0.00) 0.02(0.02+0.00) +0.0%
5311.48: server (4 days, --path-walk) 1.44(1.39+0.04) 0.01(0.01+0.00) -99.3%
5311.49: size (4 days, --path-walk) 238.1K 238.1K +0.0%
5311.50: client (4 days, --path-walk) 0.03(0.03+0.00) 0.03(0.03+0.00) +0.0%
5311.52: server (8 days, --path-walk) 1.43(1.39+0.03) 0.01(0.00+0.00) -99.3%
5311.53: size (8 days, --path-walk) 344.9K 344.9K +0.0%
5311.54: client (8 days, --path-walk) 0.07(0.07+0.00) 0.07(0.08+0.00) +0.0%
5311.56: server (16 days, --path-walk) 1.47(1.44+0.03) 0.10(0.08+0.01) -93.2%
5311.57: size (16 days, --path-walk) 844.0K 844.0K +0.0%
5311.58: client (16 days, --path-walk) 0.09(0.09+0.00) 0.09(0.09+0.00) +0.0%
5311.60: server (32 days, --path-walk) 1.52(1.50+0.05) 0.14(0.15+0.02) -90.8%
5311.61: size (32 days, --path-walk) 4.2M 4.2M +0.1%
5311.62: client (32 days, --path-walk) 0.34(0.48+0.02) 0.34(0.45+0.05) +0.0%
5311.64: server (64 days, --path-walk) 1.55(1.52+0.06) 0.15(0.15+0.04) -90.3%
5311.65: size (64 days, --path-walk) 6.4M 6.4M -0.0%
5311.66: client (64 days, --path-walk) 0.51(0.79+0.05) 0.51(0.80+0.06) +0.0%
5311.68: server (128 days, --path-walk) 1.59(1.57+0.06) 0.16(0.21+0.01) -89.9%
5311.69: size (128 days, --path-walk) 8.4M 8.4M -0.0%
5311.70: client (128 days, --path-walk) 0.72(1.44+0.08) 0.71(1.47+0.09) -1.4%
We get the same size of output pack, but this commit allows us to do so
in a significantly shorter amount of time. Intuitively, we're generating
the same pack (hence the unchanged 'test_size' output from run to run),
but varying how we get there. Before this commit, pack-objects prefers
'--path-walk' to '--use-bitmap-index', so we generate the output pack by
performing a normal '--path-walk' traversal. With this commit, we are
operating over a *repacked* state (that itself was done with a
'--path-walk' traversal), but are able to perform pack-reuse on that
repacked state via bitmaps.
When comparing the size of the repacked pack with/without '--path-walk'
on the previous commit versus this one, we see that (a) the repacked size
improves significantly with '--path-walk', and that (b) writing bitmaps
during repacking does not regress this improvement:
Test HEAD^ HEAD
----------------------------------------------------------------------------------------
5311.3: size of bitmapped pack 558.4M 558.5M +0.0%
5311.38: size of bitmapped pack (--path-walk) 164.4M 164.4M +0.0%
(Note that to observe an improvement here, we must repack with '-F' in
order to avoid reusing non-'--path-walk' deltas, which would otherwise
skew our results.)
There is one wrinkle when it comes to '--boundary', which we must not
pass into the bitmap walk in the presence of both '--path-walk' and
'--use-bitmap-index'. Path-walk needs boundary commits when it performs
its own traversal, in order to discover bases for thin packs, but the
bitmap traversal does not expect this. Work around this by setting
`revs->boundary` as late as possible within the '--path-walk' traversal,
after any bitmap attempt has either succeeded or declined to answer the
request.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
p5311 measures the cost of serving a fetch from a bitmapped pack and
indexing the resulting pack on the client. Since 761416ef91
(bitmap-lookup-table: add performance tests for lookup table,
2022-08-14), p5311 effectively runs itself twice: once with the bitmap's
lookup table extension enabled, and again with it disabled.
This comparison has served its useful purpose, as the lookup table is
almost four years old, and the de-facto default in server-side Git
deployments.
A following commit will want to test a different combination (repacking
with and without '--path-walk' instead of the lookup table). Instead of
multiplying the current test count by two again to produce four
variations of `test_fetch_bitmaps()`, drop the lookup table option to
reduce the number of perf tests we run. Retain `test_fetch_bitmaps()`
itself, since we will use this in the future for the new
parameterization.
(As an aside, a future commit outside of this series will adjust the
default value of 'pack.writeBitmapLookupTable' to "true", matching the
de-facto norm for deployments where the existence of bitmap lookup
tables is meaningful. Punt on that to a later series and instead make
the minimal change for now.)
Suggested-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
`git ls-files --modified` and `git ls-files --deleted` have been
optimized to filter with pathspec before calling lstat() when there is
only a single pathspec item, avoiding unnecessary filesystem access
for entries that will not be shown.
* td/ls-files-pathspec-prefilter:
ls-files: filter pathspec before lstat
'git describe' has been taught to pass the 'refs/tags/' prefix down to
the ref iterator when '--all' is not requested, avoiding unnecessary
iteration over non-tag refs.
* td/describe-tag-iteration:
describe: limit default ref iteration to tags
git branch and git for-each-ref run a separate reachability walk for
each ref considered by --contains and --no-contains. Refs with shared
history therefore traverse the same commits repeatedly.
git tag instead uses a depth-first walk that caches results across
refs. That walk can perform poorly without generation numbers: a
negative check may walk to the root instead of stopping at a nearby
divergence. Generation numbers let it stop below the oldest target.
Use the memoized walk for all ref-filter callers when generation
numbers are available. Keep git tag on its existing path without
generations. Caching still helps when many tags share deep history:
ffc4b8012d (tag: speed up --contains calculation, 2011-06-11) reduced
git tag --contains HEAD~200 in linux-2.6 from 15.417 to 5.329 seconds.
The new shared-history perf test improves from 0.72 to 0.03 seconds. In
a repository with 62,174 remote-tracking refs, running:
git branch -r --contains c78ae85f3ce7e
improves from 104.365 seconds to 468 milliseconds.
Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Tamir Duberstein <tamird@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
dabecb9db2 (for-each-ref: introduce a '--start-after' option,
2025-07-15) changed branch, remote-tracking branch, and tag enumeration
from constructing an iterator with the namespace prefix to constructing
an unscoped iterator and seeking to the prefix.
Review of --start-after noted that the construction prefix and seek
position represent different state and are easy to conflate [1]. It also
noted that future branch or tag support would need to retain the
namespace prefix while moving the cursor [2].
The files backend constructs its loose-ref iterator with cache priming
enabled. cache_ref_iterator_begin() immediately applies the construction
prefix through cache_ref_iterator_set_prefix(), reading loose refs
beneath it before packed refs are opened. An empty prefix therefore
reads every loose ref, and a later seek cannot undo that I/O.
For the current single-kind filters, construct the iterator with the
namespace prefix when start_after is not set. Leave the existing
start_after path unchanged; no current command combines it with these
filters, and future support must carry the prefix separately from the
cursor.
With 10,000 unrelated loose refs in the files backend, the p6300 tests
improve as follows:
before after
branch 2.74 s 0.11 s
branch --remotes 2.81 s 0.12 s
tag 3.01 s 0.11 s
[1] https://lore.kernel.org/r/aGZidwwlToWThkn8@pks.im/
[2] https://lore.kernel.org/r/xmqqikjq7s16.fsf@gitster.g/
Fixes: dabecb9db2 ("for-each-ref: introduce a '--start-after' option")
Suggested-by: Karthik Nayak <karthik.188@gmail.com>
Signed-off-by: Tamir Duberstein <tamird@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In --deleted and --modified modes, show_files() calls lstat() for each
index entry before show_ce() applies the pathspec. prune_index() avoids
most of these calls for pathspecs with a common directory prefix, but
not for a top-level name or leading wildcard.
Match before lstat() to avoid accessing the worktree for entries that
cannot be shown. Treat this as a prefilter: do not update ps_matched,
and retain the match in show_ce() so --error-unmatch is satisfied only
by entries that the selected modes actually show.
Prefilter only a single pathspec item, bounding the added work for each
index entry. Applying match_pathspec() to multiple arguments can cost
more than the lstat() calls it avoids. In a synthetic repository with
10,000 clean files, passing every path to ls-files --modified increased
runtime from 112.5 ms to 494.1 ms when the prefilter was unconditional.
With $parent and $this exported as paths to binaries built from the
parent and this commit, on a repository with 881,290 index entries:
hyperfine --warmup 0 --runs 3 \
--command-name parent \
'$parent -c core.fsmonitor=false ls-files --deleted -- README.md >/dev/null' \
--command-name this-commit \
'$this -c core.fsmonitor=false ls-files --deleted -- README.md >/dev/null'
reported means of 65.790 seconds for the parent and 4.987 seconds for
this commit.
Link: https://lore.kernel.org/r/xmqqfr2tnfk0.fsf@gitster.g
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Tamir Duberstein <tamird@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Without --all, git describe ignores refs outside refs/tags/. Commit
8a5a1884e9 (Avoid accessing non-tag refs in git-describe unless --all is
requested, 2008-02-24) moved this check ahead of object lookup. That
avoided loading objects for irrelevant refs, but the backend still has
to yield every ref before get_name() can reject it.
Pass refs/tags/ to the iterator so the backend can avoid visiting those
refs in the first place.
The new perf test creates 10,000 unrelated packed refs. It measures:
git describe --exact-match HEAD
The runtime drops from 0.03(0.01+0.01) to 0.02(0.00+0.00). In a
repository with 120,532 refs but only 330 tags, the same command went
from 171.7 ms to 9.9 ms.
Signed-off-by: Tamir Duberstein <tamird@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git stash -p" has been optimized by reusing cached index
entries in its temporary index, avoiding unnecessary lstat()
calls on unchanged files.
* aj/stash-patch-optimize-temporary-index:
stash: reuse cached index entries in --patch temporary index
`git stash -p` prepares the interactive selection by creating a
temporary index at HEAD, switching `GIT_INDEX_FILE` to it, and then
running the `add -p` machinery.
That temporary index was created by running `git read-tree HEAD`. The
resulting index had no useful cached stat data or fsmonitor-valid bits
from the real index. When `run_add_p()` refreshed that temporary index
before showing the first prompt, it could end up lstat(2)-ing every
tracked file, even in a repository where `git diff` and `git restore -p`
can use fsmonitor to avoid that work.
Create the temporary index in-process instead. Use `unpack_trees()` to
reset the real index contents to HEAD while writing the result to the
temporary index path. For paths whose index entries already match HEAD,
`oneway_merge()` reuses the existing cache entries, preserving their
cached stat data and `CE_FSMONITOR_VALID` state.
This makes the refresh performed by `run_add_p()` behave like the one
used by `git restore -p`: unchanged paths can be skipped via fsmonitor
instead of being scanned again.
In a 206k file repository with `core.fsmonitor` enabled and a one-line
change in one file, time to first prompt dropped from 34.774 seconds to
0.659 seconds. The new perf test file demonstrates similar improvements,
with maen times for without- and with-fsmonitor cases dropping from 6.90
and 6.83 seconds to 0.55 and 0.28 seconds, respectively.
Signed-off-by: Adam Johnson <me@adamj.eu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add p5315-pack-objects-filter.sh to measure the performance of
'git pack-objects --revs --all' under different filter and traversal
combinations:
* no filter (baseline)
* --filter=blob:none (blobless)
* --filter=sparse:oid=<oid> (cone-mode sparse)
Each filter scenario is tested both with and without --path-walk,
producing paired measurements that show the impact of the path-walk
traversal for each filter type as we integrate the --path-walk feature
with different --filter options. It currently has no integration so
falls back to the standard revision walk. Thus, there are no significant
differences in the current results other than a full repack (and even
then, the --path-walk feature is not incredibly different for the
default Git repository):
Test HEAD
-----------------------------------------------------
5315.2: repack (no filter) 27.91
5315.3: repack size (no filter) 250.7M
5315.4: repack (no filter, --path-walk) 34.92
5315.5: repack size (no filter, --path-walk) 220.0M
5315.6: repack (blob:none) 13.63
5315.7: repack size (blob:none) 137.6M
5315.8: repack (blob:none, --path-walk) 13.48
5315.9: repack size (blob:none, --path-walk) 137.7M
5315.10: repack (sparse:oid) 72.67
5315.11: repack size (sparse:oid) 187.4M
5315.12: repack (sparse:oid, --path-walk) 72.47
5315.13: repack size (sparse:oid, --path-walk) 187.4M
The sparse filter definition is built automatically by sampling
depth-2 directories from the test repository, making the test work
on any repo passed via GIT_PERF_LARGE_REPO. For repos that lack
depth-2 directories, a single top-level directory is used; for flat
repos, the sparse tests are skipped via prerequisite.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a performance test that compares 'git rev-list --maximal-only'
against 'git merge-base --independent'. These two commands are asking
essentially the same thing, but the rev-list implementation is more
generic and hence slower. These performance tests will demonstrate that
in the current state and also be used to show the equivalence in the
future.
We also add a case with '--since' to force the generic walk logic for
rev-list even when we make that future change to use the merge-base
algorithm on a simple walk.
When run on my copy of git.git, I see these results:
Test HEAD
----------------------------------------------
6011.2: merge-base --independent 0.03
6011.3: rev-list --maximal-only 0.06
6011.4: rev-list --maximal-only --since 0.06
These numbers are low, but the --independent calculation is interesting
due to having a lot of local branches that are actually independent.
Running the same test on a fresh clone of the Linux kernel repository
shows a larger difference between the algorithms, especially because the
--independent algorithm is extremely fast when there are no independent
references selected:
Test HEAD
----------------------------------------------
6011.2: merge-base --independent 0.00
6011.3: rev-list --maximal-only 0.70
6011.4: rev-list --maximal-only --since 0.70
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The setup phase in 't/perf/p3400-rebase.sh' generates 100 commits to
simulate a noisy history. It currently uses a shell loop that invokes
'git add', 'git commit', 'test_seq', and 'sort' in each iteration.
This incurs significant overhead due to repeated process spawning.
Optimize the setup by using 'git fast-import' to generate the commit
history. Additionally, pre-compute the forward and reversed file contents
to avoid repetitive execution of 'seq' and 'sort'.
To ensure the test measures rebase performance against a consistent
object layout (rather than the suboptimal pack/loose objects created
by the raw import), perform a full repack (`git repack -a -d`) at the
end of the setup.
This reduces the setup time significantly while maintaining the validity
of the subsequent performance tests.
Performance enhancement (Average value of 5 tests):
Real Rebase
Before: 29.045s 13.34s
After: 21.989s 12.84s
Measured on Lenovo Yoga 2020, Ubuntu 24.04.
Signed-off-by: Tian Yuchen <a3205153416@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 8002e8ee18 (builtin/cat-file: use bitmaps to efficiently filter
by object type, 2025-04-02) introduced a performance regression when we
are not filtering objects: it uses bitmaps even when they won't help,
incurring extra costs. For example, running the new perf tests from this
commit, which check the performance of listing objects by oid:
$ export GIT_PERF_LARGE_REPO=/path/to/linux.git
$ git -C "$GIT_PERF_LARGE_REPO" repack -adb
$ GIT_SKIP_TESTS=p1006.1 ./run 8002e8ee18^ 8002e8ee18 p1006-cat-file.sh
[...]
Test 8002e8ee18^ 8002e8ee18
-------------------------------------------------------------------------------
1006.2: list all objects (sorted) 1.48(1.44+0.04) 6.39(6.35+0.04) +331.8%
1006.3: list all objects (unsorted) 3.01(2.97+0.04) 3.40(3.29+0.10) +13.0%
1006.4: list blobs 4.85(4.67+0.17) 1.68(1.58+0.10) -65.4%
An invocation that filters, like listing all blobs (1006.4), does
benefit from using the bitmaps; it now doesn't have to check the type of
each object from the pack data, so the tradeoff is worth it.
But for listing all objects in sorted idx order (1006.2), we otherwise
would never open the bitmap nor the revindex file. Worse, our sorting
step gets much worse. Normally we append into an array in pack .idx
order, and the sort step is trivial. But with bitmaps, we get the
objects in pack order, which is apparently random with respect to oid,
and have to sort the whole thing. (Note that this freshly-packed state
represents the best case for .idx sorting; if we had two packs, then
we'd have their objects one after the other and qsort would have to
interleave them).
The unsorted test in 1006.3 is interesting: there we are going in pack
order, so we load the revindex for the pack anyway. And though we don't
sort the result, we do use an oidset to check for duplicates. So we can
see in the 8002e8ee18^ timings that those two things cost ~1.5s over the
sorted case (mostly the oidset hash cost). We also incur the extra cost
to open the bitmap file as of 8002e8ee18, which seems to be ~400ms.
(This would probably be faster with a bitmap lookup table, but writing
that out is not yet the default).
So we know that bitmaps help when there's filtering to be done, but
otherwise make things worse. Let's only use them when there's a filter.
The perf script shows that we've fixed the regressions without hurting
the bitmap case:
Test 8002e8ee18^ 8002e8ee18 HEAD
--------------------------------------------------------------------------------------------------------
1006.2: list all objects (sorted) 1.56(1.53+0.03) 6.44(6.37+0.06) +312.8% 1.62(1.54+0.06) +3.8%
1006.3: list all objects (unsorted) 3.04(2.98+0.06) 3.45(3.38+0.07) +13.5% 3.04(2.99+0.04) +0.0%
1006.4: list blobs 5.14(4.98+0.15) 1.76(1.68+0.06) -65.8% 1.73(1.64+0.09) -66.3%
Note that there's another related case: we might have a filter that
cannot be used with bitmaps. That check is handled already for us in
for_each_bitmapped_object(), though we'd still load the bitmap and
revindex files pointlessly in that case. I don't think it can happen in
practice for cat-file, though, since it allows only blob:none,
blob:limit, and object:type filters, all of which work with bitmaps.
It would be easy-ish to insert an extra check like:
can_filter_bitmap(&opt->objects_filter);
into the conditional, but I didn't bother here. It would be redundant
with the call in for_each_bitmapped_object(), and the can_filter helper
function is static local in the bitmap code (so we'd have to make it
public).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If you run:
GIT_PERF_LARGE_REPO=/some/path ./p1006-cat-file.sh
it will use the repo in /some/path. But if you use the "run" helper
script to aggregate and compare results, like this:
GIT_PERF_LARGE_REPO=/some/path ./run HEAD^ HEAD p1006-cat-file.sh
it will ignore that variable. This is because the presence of the
LARGE_REPO variable in GIT-BUILD-OPTIONS overrides what's in the
environment. This started with 4638e8806e (Makefile: use common template
for GIT-BUILD-OPTIONS, 2024-12-06), which now writes even empty
variables (though arguably it was wrong even before with a non-empty
value, as we generally prefer the environment to take precedence over
on-disk config).
We had the same problem in perf-lib.sh itself, and we hacked around it
with 32b74b9809 (perf: do allow `GIT_PERF_*` to be overridden again,
2025-04-04). That's what lets the direct invocation of "./p1006" work
above.
And in fact that was sufficient for "./run", too, until it started
loading GIT-BUILD-OPTIONS itself in 5756ccd181 (t/perf: fix benchmarks
with out-of-tree builds, 2025-04-28). Now it has the same problem: it
clobbers any incoming GIT_PERF options from the environment.
We can use the same hack here in the "run" script. It's quite ugly, but
it's just short enough that I don't think it's worth trying to factor it
out into a common shell library.
In the long run, we might consider teaching GIT-BUILD-OPTIONS to be more
gentle in overwriting existing entries. There are probably other
GIT_TEST_* variables which would need the same treatment. And if and
when we come up with a more complete solution, we can use it in both
spots.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Using the perf suite's "run" helper in a vanilla build fails like this:
$ make && (cd t/perf && ./run p0000-perf-lib-sanity.sh)
=== Running 1 tests in this tree ===
perf 1 - test_perf_default_repo works: 1 2 3 ok
perf 2 - test_checkout_worktree works: 1 2 3 ok
ok 3 - test_export works
perf 4 - export a weird var: 1 2 3 ok
perf 5 - éḿíẗ ńöń-ÁŚĆÍÍ ćḧáŕáćẗéŕś: 1 2 3 ok
ok 6 - test_export works with weird vars
perf 7 - important variables available in subshells: 1 2 3 ok
perf 8 - test-lib-functions correctly loaded in subshells: 1 2 3 ok
# passed all 8 test(s)
1..8
cannot open test-results/p0000-perf-lib-sanity.subtests: No such file or directory at ./aggregate.perl line 159.
It is trying to aggregate results written into t/perf/test-results, but
the p0000 script did not write anything there.
The "run" script looks in $TEST_OUTPUT_DIRECTORY/test-results, or if
that variable is not set, in test-results in the current working
directory (which should be t/perf itself). It pulls the value of
$TEST_OUTPUT_DIRECTORY from the GIT-BUILD-OPTIONS file.
But that doesn't quite match the setup in perf-lib.sh (which is what
scripts like p0000 use). There we do this at the top of the script:
TEST_OUTPUT_DIRECTORY=$(pwd)
and then let test-lib.sh append "/test-results" to that. Historically,
that made the vanilla case work: we'd always use t/perf/test-results.
But when $TEST_OUTPUT_DIRECTORY was set, it would break.
Commit 5756ccd181 (t/perf: fix benchmarks with out-of-tree builds,
2025-04-28) fixed that second case by loading GIT-BUILD-OPTIONS
ourselves. But that broke the vanilla case!
Now our setting of $TEST_OUTPUT_DIRECTORY in perf-lib.sh is ignored,
because it is overwritten by GIT-BUILD-OPTIONS. And when test-lib.sh
sees that the output directory is empty, it defaults to t/test-results,
rather than t/perf/test-results.
Nobody seems to have noticed, probably for two reasons:
1. It only matters if you're trying to aggregate results (like the
"run" script does). Just running "./p0000-perf-lib-sanity.sh"
manually still produces useful output; the stored result files are
just in an unexpected place.
2. There might be leftover files in t/perf/test-results from previous
runs (before 5756ccd181). In particular, the ".subtests" files
don't tend to change, and the lack of that file is what causes it
to barf completely. So it's possible that the aggregation could
have been showing stale results that did not match the run that
just happened.
We can fix it by setting TEST_OUTPUT_DIRECTORY only after we've loaded
GIT-BUILD-OPTIONS, so that we override its value and not the other way
around. And we'll do so only when the variable is not set, which should
retain the fix for that case from 5756ccd181.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Building a list using commit_list_insert_by_date() has quadratic worst
case complexity. Avoid it by using prio_queue.
Use prio_queue_peek()+prio_queue_replace() instead of prio_queue_get()+
prio_queue_put() if possible, as the former only rebalance the
prio_queue heap once instead of twice.
In sane repositories this won't make much of a difference because the
number of items in the list or queue won't be very high:
Benchmark 1: ./git_v2.52.0 show-branch origin/main origin/next origin/seen origin/todo
Time (mean ± σ): 538.2 ms ± 0.8 ms [User: 527.6 ms, System: 9.6 ms]
Range (min … max): 537.0 ms … 539.2 ms 10 runs
Benchmark 2: ./git show-branch origin/main origin/next origin/seen origin/todo
Time (mean ± σ): 530.6 ms ± 0.4 ms [User: 519.8 ms, System: 9.8 ms]
Range (min … max): 530.1 ms … 531.3 ms 10 runs
Summary
./git show-branch origin/main origin/next origin/seen origin/todo ran
1.01 ± 0.00 times faster than ./git_v2.52.0 show-branch origin/main origin/next origin/seen origin/todo
That number is not limited, though, and in pathological cases like the
one in p6010 we see a sizable improvement:
Test v2.52.0 HEAD
------------------------------------------------------------------
6010.4: git show-branch 2.19(2.19+0.00) 0.03(0.02+0.00) -98.6%
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Building a list using commit_list_insert_by_date() has quadratic worst
case complexity. Avoid it by just appending in the loop and sorting at
the end.
The number of merge bases is usually small, so don't expect speedups in
normal repositories. It has no limit, though. The added perf test
shows a nice improvement when dealing with 16384 merge bases:
Test v2.51.1 HEAD
-----------------------------------------------------------------
6010.2: git merge-base 0.55(0.54+0.00) 0.03(0.02+0.00) -94.5%
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This just runs some simple last-modified commands. We already test
correctness in the regular suite, so this is just about finding
performance regressions from one version to another.
Based-on-patch-by: Jeff King <peff@peff.net>
Signed-off-by: Toon Claes <toon@iotcl.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The pop_most_recent_commit() function can have quite expensive
worst case performance characteristics, which has been optimized by
using prio-queue data structure.
* rs/pop-recent-commit-with-prio-queue:
commit: use prio_queue_replace() in pop_most_recent_commit()
prio-queue: add prio_queue_replace()
commit: convert pop_most_recent_commit() to prio_queue
pop_most_recent_commit() calls commit_list_insert_by_date() for parent
commits, which is itself called in a loop. This can lead to quadratic
complexity if there are many merges. Replace the commit_list with a
prio_queue to ensure logarithmic worst case complexity and convert all
three users.
Add a performance test that exercises one of them using a pathological
history that consists of 50% merges and 50% root commits to demonstrate
the speedup:
Test v2.50.1 HEAD
----------------------------------------------------------------------
1501.2: rev-parse ':/65535' 2.48(2.47+0.00) 0.20(0.19+0.00) -91.9%
Alas, sane histories don't benefit from the conversion much, and
traversing Git's own history takes a 1% performance hit on my machine:
$ hyperfine -w3 -L git ./git_2.50.1,./git '{git} rev-parse :/^Initial.revision'
Benchmark 1: ./git_2.50.1 rev-parse :/^Initial.revision
Time (mean ± σ): 1.071 s ± 0.004 s [User: 1.052 s, System: 0.017 s]
Range (min … max): 1.067 s … 1.078 s 10 runs
Benchmark 2: ./git rev-parse :/^Initial.revision
Time (mean ± σ): 1.079 s ± 0.003 s [User: 1.060 s, System: 0.017 s]
Range (min … max): 1.074 s … 1.083 s 10 runs
Summary
./git_2.50.1 rev-parse :/^Initial.revision ran
1.01 ± 0.00 times faster than ./git rev-parse :/^Initial.revision
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git pack-objects" learns to find delta bases from blobs at the
same path, using the --path-walk API.
* ds/path-walk-2:
pack-objects: allow --shallow and --path-walk
path-walk: add new 'edge_aggressive' option
pack-objects: thread the path-based compression
pack-objects: refactor path-walk delta phase
scalar: enable path-walk during push via config
pack-objects: enable --path-walk via config
repack: add --path-walk option
t5538: add tests to confirm deltas in shallow pushes
pack-objects: introduce GIT_TEST_PACK_PATH_WALK
p5313: add performance tests for --path-walk
pack-objects: update usage to match docs
pack-objects: add --path-walk option
pack-objects: extract should_attempt_deltas()
Remove the leftover hints to the test framework to mark tests that
do not pass the leak checker tests, as they should no longer be
needed.
* kn/passing-leak-tests:
t: remove unexpected SANITIZE_LEAK variables
"git apply" and "git add -i/-p" code paths no longer unnecessarily
expand sparse-index while working.
* ds/sparse-apply-add-p:
p2000: add performance test for patch-mode commands
reset: integrate sparse index with --patch
git add: make -p/-i aware of sparse index
apply: integrate with the sparse index
As of 1fc7ddf35b (test-lib: unconditionally enable leak checking,
2024-11-20), both the `GIT_TEST_PASSING_SANITIZE_LEAK` and
`TEST_PASSES_SANITIZE_LEAK` variables no longer have any meaning, the
leak checks are enabled by default. However, some newly added tests
include them by mistake. Let's clean this up.
Signed-off-by: Karthik Nayak <karthik.188@gmail.com>
Acked-by: Justin Tobler <jltobler@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since 'git pack-objects' supports a --path-walk option, allow passing it
through in 'git repack'. This presents interesting testing opportunities for
comparing the different repacking strategies against each other.
Add the --path-walk option to the performance tests in p5313.
For the microsoft/fluentui repo [1] checked out at a specific commit [2],
the --path-walk tests in p5313 look like this:
Test this tree
-------------------------------------------------------------------------
5313.18: thin pack with --path-walk 0.08(0.06+0.02)
5313.19: thin pack size with --path-walk 18.4K
5313.20: big pack with --path-walk 2.10(7.80+0.26)
5313.21: big pack size with --path-walk 19.8M
5313.22: shallow fetch pack with --path-walk 1.62(3.38+0.17)
5313.23: shallow pack size with --path-walk 33.6M
5313.24: repack with --path-walk 81.29(96.08+0.71)
5313.25: repack size with --path-walk 142.5M
[1] https://github.com/microsoft/fluentui
[2] e70848ebac1cd720875bccaa3026f4a9ed700e08
Along with the earlier tests in p5313, I'll instead reformat the
comparison as follows:
Repack Method Pack Size Time
---------------------------------------
Hash v1 439.4M 87.24s
Hash v2 161.7M 21.51s
Path Walk 142.5M 81.29s
There are a few things to notice here:
1. The benefits of --name-hash-version=2 over --name-hash-version=1 are
significant, but --path-walk still compresses better than that
option.
2. The --path-walk command is still using --name-hash-version=1 for the
second pass of delta computation, using the increased name hash
collisions as a potential method for opportunistic compression on
top of the path-focused compression.
3. The --path-walk algorithm is currently sequential and does not use
multiple threads for delta compression. Threading will be
implemented in a future change so the computation time will improve
to better compete in this metric.
There are small benefits in size for my copy of the Git repository:
Repack Method Pack Size Time
---------------------------------------
Hash v1 248.8M 30.44s
Hash v2 249.0M 30.15s
Path Walk 213.2M 142.50s
As well as in the nodejs/node repository [3]:
Repack Method Pack Size Time
---------------------------------------
Hash v1 739.9M 71.18s
Hash v2 764.6M 67.82s
Path Walk 698.1M 208.10s
[3] https://github.com/nodejs/node
This benefit also repeats in my copy of the Linux kernel repository:
Repack Method Pack Size Time
---------------------------------------
Hash v1 2.5G 554.41s
Hash v2 2.5G 549.62s
Path Walk 2.2G 1562.36s
It is important to see that even when the repository shape does not have
many name-hash collisions, there is a slight space boost to be found
using this method.
As this repacking strategy was released in Git for Windows 2.47.0, some
users have reported cases where the --path-walk compression is slightly
worse than the --name-hash-version=2 option. In those cases, it may be
beneficial to combine the two options. However, there has not been a
released version of Git that has both options and I don't have access to
these repos for testing.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous change added a --path-walk option to 'git pack-objects'.
Create a performance test that demonstrates the time and space benefits
of the feature.
In order to get an appropriate comparison, we need to avoid reusing
deltas and recompute them from scratch.
Compare the creation of a thin pack representing a small push and the
creation of a relatively large non-thin pack.
Running on my copy of the Git repository results in this data (removing
the repack tests for --name-hash-version):
Test this tree
------------------------------------------------------------------------
5313.2: thin pack with --name-hash-version=1 0.02(0.01+0.01)
5313.3: thin pack size with --name-hash-version=1 1.6K
5313.4: big pack with --name-hash-version=1 2.55(4.20+0.26)
5313.5: big pack size with --name-hash-version=1 16.4M
5313.6: shallow fetch pack with --name-hash-version=1 1.24(2.03+0.08)
5313.7: shallow pack size with --name-hash-version=1 12.2M
5313.10: thin pack with --name-hash-version=2 0.03(0.01+0.01)
5313.11: thin pack size with --name-hash-version=2 1.6K
5313.12: big pack with --name-hash-version=2 1.91(3.23+0.20)
5313.13: big pack size with --name-hash-version=2 16.4M
5313.14: shallow fetch pack with --name-hash-version=2 1.06(1.57+0.10)
5313.15: shallow pack size with --name-hash-version=2 12.5M
5313.18: thin pack with --path-walk 0.03(0.01+0.01)
5313.19: thin pack size with --path-walk 1.6K
5313.20: big pack with --path-walk 2.05(3.24+0.27)
5313.21: big pack size with --path-walk 16.3M
5313.22: shallow fetch pack with --path-walk 1.08(1.66+0.07)
5313.23: shallow pack size with --path-walk 12.4M
This can be reformatted as follows:
Pack Type Hash v1 Hash v2 Path Walk
---------------------------------------------------
thin pack (time) 0.02s 0.03s 0.03s
(size) 1.6K 1.6K 1.6K
big pack (time) 2.55s 1.91s 2.05s
(size) 16.4M 16.4M 16.3M
shallow pack (time) 1.24s 1.06s 1.08s
(size) 12.2M 12.5M 12.4M
Note that the timing is slower because there is no threading in the
--path-walk case (yet). Also, the shallow pack cases are really not
using the --path-walk logic right now because it is disabled until some
additions are made to the path walk API.
The cases where the --path-walk option really shines is when the default
name-hash is overwhelmed with unhelpful collisions. An open source
example can be found in the microsoft/fluentui repo [1] at a certain
commit [2].
[1] https://github.com/microsoft/fluentui
[2] e70848ebac1cd720875bccaa3026f4a9ed700e08
Running the tests on this repo results in the following comparison table:
Pack Type Hash v1 Hash v2 Path Walk
---------------------------------------------------
thin pack (time) 0.36s 0.12s 0.08s
(size) 1.2M 22.0K 18.4K
big pack (time) 2.00s 2.90s 2.21s
(size) 20.4M 25.9M 19.5M
shallow pack (time) 1.41s 1.80s 1.65s
(size) 34.4M 33.7M 33.6M
Notice in particular that in the small thin pack, the time performance
has improved from 0.36s for --name-hash-version=1 to 0.08s and this is
likely due to the improved size of the resulting pack: 18.4K instead of
1.2M. The relatively new --name-hash-version=2 is competitive with
--path-walk (0.12s and 22.0K) but not quite as successful.
Finally, running this on a copy of the Linux kernel repository results
in these data points:
Pack Type Hash v1 Hash v2 Path Walk
---------------------------------------------------
thin pack (time) 0.03s 0.13s 0.03s
(size) 4.6K 4.6K 4.6K
big pack (time) 15.29s 12.32s 13.92s
(size) 201.1M 159.1M 158.5M
shallow pack (time) 10.88s 22.93s 22.74s
(size) 269.2M 273.8M 267.7M
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous three changes contributed performance improvements to 'git
apply', 'git add -p', and 'git reset -p' when using a sparse index. The
improvement to 'git apply' also improved 'git checkout -p'. Add
performance tests to demonstrate this (and to help validate that
performance remains good in the future).
In the truncated test output below, we see that the full checkout
performance changes within noise expectations, but the sparse index
cases improve 33% and then 96% for 'git add -p' and 41% and then 95% for
'git reset -p'. 'git checkout -p' improves immediatley by 91% because it
does not need any change to its builtin.
Test HEAD~4 HEAD~3 HEAD~2 HEAD~1
-------------------------------------------------------------------------------------
2000.118: ... git add -p (full-v3) 0.79 0.79 +0.0% 0.82 +3.8% 0.82 +3.8%
2000.119: ... git add -p (full-v4) 0.74 0.76 +2.7% 0.74 +0.0% 0.76 +2.7%
2000.120: ... git add -p (sparse-v3) 1.94 1.28 -34.0% 0.07 -96.4% 0.07 -96.4%
2000.121: ... git add -p (sparse-v4) 1.93 1.28 -33.7% 0.06 -96.9% 0.06 -96.9%
2000.122: ... git checkout -p (full-v3) 1.18 1.18 +0.0% 1.18 +0.0% 1.19 +0.8%
2000.123: ... git checkout -p (full-v4) 1.10 1.12 +1.8% 1.11 +0.9% 1.11 +0.9%
2000.124: ... git checkout -p (sparse-v3) 1.31 0.11 -91.6% 0.11 -91.6% 0.11 -91.6%
2000.125: ... git checkout -p (sparse-v4) 1.29 0.11 -91.5% 0.11 -91.5% 0.11 -91.5%
2000.126: ... git reset -p (full-v3) 0.81 0.80 -1.2% 0.83 +2.5% 0.83 +2.5%
2000.127: ... git reset -p (full-v4) 0.78 0.77 -1.3% 0.77 -1.3% 0.78 +0.0%
2000.128: ... git reset -p (sparse-v3) 1.58 0.92 -41.8% 0.91 -42.4% 0.07 -95.6%
2000.129: ... git reset -p (sparse-v4) 1.58 0.92 -41.8% 0.92 -41.8% 0.07 -95.6%
It is worth noting that if our test was more involved and had multiple
hunks to evaluate, then the time spent in 'git apply' would dominate due
to multiple index loads and writes. As it stands, we need the sparse
index improvement in 'git add -p' itself to confirm this performance
improvement.
Since the change for 'git add -i' is identical, we avoid a second test
case for that similar operation.
Signed-off-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The build procedure based on Meson learned to drive the
benchmarking tests.
* ps/meson-build-perf-bench:
meson: wire up benchmarking options
meson: wire up benchmarks
t/perf: fix benchmarks with out-of-tree builds
t/perf: use configured PERL_PATH
t/perf: fix benchmarks with alternate repo formats
The "perf-lib.sh" script is sourced by all of our benchmarking suites to
make available common infrastructure. The script assumes that build and
source directory are the same, which works for our Makefile. But the
assumption breaks with both CMake and Meson, where the build directory
can be located in an arbitrary place.
Adapt the script so that it works with out-of-tree builds. Most
importantly, this requires us to figure out the location of the build
directory:
- When running benchmarks via our Makefile the build directory is the
same as the source directory. We already know to derive the test
directory ("t/") via `$(pwd)/..`, which works because we chdir into
"t/perf" before executing benchmarks. We can thus derive the build
directory by appending another "/.." to that path.
- When running benchmarks via Meson the build directory is located at
an arbitrary location. The build system thus has to make the path
known by exporting the `GIT_BUILD_DIR` environment variable.
This change prepares us for wiring up benchmarks in Meson.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our benchmarks use a couple of Perl scripts to compute results. These
Perl scripts get executed directly, and as the shebang is hardcoded to
"/usr/bin/perl" this will fail on any system where the Perl interpreter
is located in a different path.
Our build infrastructure already lets users configure the location of
Perl, which ultimately gets written into the GIT-BUILD-OPTIONS file.
This file is being sourced by "test-lib.sh", and consequently we already
have the "PERL_PATH" variable available that contains its configured
location.
Use "PERL_PATH" to execute Perl scripts, which makes them work on more
esoteric systems like NixOS. Furthermore, adapt the shebang to use
env(1) to execute Perl so that users who have Perl in PATH, but in a
non-standard location can execute the script directly.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Many of our benchmarks operate on a user-defined repository that we copy
over before running the benchmarked logic. To keep unintentional side
effects caused by on-disk state at bay we skip copying some files. This
includes for example hooks, but also the repo's configuration.
It is quite sensible to not copy over the configuration, as it is quite
easy to inadvertently carry over configuration that may significantly
impact the performance measurements. But we cannot fully ignore the
configuration either, as it may contain information about the repository
format. This will cause failures when for example using a repository
with SHA256 object format or the reftable ref format.
Fix the issue by parsing the reference and object formats from the
source repository and passing them to git-init(1).
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This perf script creates a midx by running "git multi-pack-index write"
with the "--stdin-packs" option. We feed that stdin by running "find" on
.git/objects/pack, using sed to strip off everything but the basename.
But that sed invocation also does something peculiar: it adds a "+" to
the start of each pack name. This causes the multi-pack-index command to
barf. The modified name does not match any pack it knows about, so it
ends up with an empty list of packs to put in the midx. And thus nothing
matches the --preferred-pack option we pass, which causes it die().
The fix is to remove the extra "+" (which also lets us simplify the sed
invocation a bit, as it is now just stripping the leading directories).
But that leaves the mystery of why it was ever there in the first place.
The answer is that an earlier iteration of the patch series had a
concept of "disjoint" packs in the midx. And one of its patches here:
https://lore.kernel.org/git/c52d7e7b27a9add4f58b8334db4fe4498af1c90f.1701198172.git.me@ttaylorr.com/
taught read_packs_from_stdin() to treat a leading "+" as marking a
disjoint pack. But in the second version of the series, which was
ultimately merged, that disjoint concept went away, and the code to
parse "+" did likewise. The regular regression tests were adjusted to
match, but this case in t/perf was forgotten.
Signed-off-by: Jeff King <peff@peff.net>
Acked-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A common way to run Git's performance benchmarks on repositories other
than Git's own repository (which is not exactly large when compared to
actually large repositories) is to run them like this:
GIT_PERF_LARGE_REPO=/path/to/my/large/repo \
./p1234-*.sh -ivx
Contrary to developers' common expectations, this failed to work when
Git was built with a different `GIT_PERF_LARGE_REPO` value specified at
build time: That build-time option would have been written to the
`GIT-BUILD-OPTIONS` file, which in turn would have been sourced by
`test-lib.sh`, which in turn would have been sourced by `perf-lib.sh`,
which in turn would have been sourced by the perf test script,
_overriding_ the environment variable specified in the way illustrated
above.
Since perf tests are not run as part of the build, this most likely
unintended behavior was not caught and certainly not fixed, as the
`GIT_PERF_*` values would have been empty at build-time.
However, in 4638e8806e (Makefile: use common template for
GIT-BUILD-OPTIONS, 2024-12-06), a subtle change of behavior was
introduced: Whereas before, a couple of build-time options (the
`GIT_PERF_*` ones included) were written to `GIT-BUILD-OPTIONS` only
when their values were non-empty. With this commit, they are also
written when they are empty.
The consequence is that above-mentioned way to run the perf tests will
not only fail to pick up the desired `GIT_PERF_*` settings when they
were specified differently while building Git, instead the desired
settings will be only respected when specified _while building_ Git.
Let's work around the original issue, i.e. let `GIT_PERF_*` environment
variables override what is recorded in `GIT-BUILD-OPTIONS`.
Note that this is just the tip of the iceberg, there are a couple of
`GIT_TEST_*` options that may want a similar fix in `test-lib.sh`. Due
to time constraints on my side, this here patch focuses exclusively on
the `GIT_PERF_*` settings.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Junio C Hamano