Add a test for the -G and -S pickaxe options and related options.
This test supports being run with GIT_TEST_LONG=1 to adjust the limit
on the number of commits from 1k to 10k. The 1k limit seems to hit a
good spot on git.git
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git repack -A -d" in a partial clone unnecessarily loosened
objects in promisor pack.
* rs/repack-without-loosening-promised-objects:
repack: avoid loosening promisor objects in partial clones
Builds on top of the sparse-index infrastructure to mark operations
that are not ready to mark with the sparse index, causing them to
fall back on fully-populated index that they always have worked with.
* ds/sparse-index-protections: (47 commits)
name-hash: use expand_to_path()
sparse-index: expand_to_path()
name-hash: don't add directories to name_hash
revision: ensure full index
resolve-undo: ensure full index
read-cache: ensure full index
pathspec: ensure full index
merge-recursive: ensure full index
entry: ensure full index
dir: ensure full index
update-index: ensure full index
stash: ensure full index
rm: ensure full index
merge-index: ensure full index
ls-files: ensure full index
grep: ensure full index
fsck: ensure full index
difftool: ensure full index
commit: ensure full index
checkout: ensure full index
...
When `git repack -A -d` is run in a partial clone, `pack-objects`
is invoked twice: once to repack all promisor objects, and once to
repack all non-promisor objects. The latter `pack-objects` invocation
is with --exclude-promisor-objects and --unpack-unreachable, which
loosens all objects unused during this invocation. Unfortunately,
this includes promisor objects.
Because the -d argument to `git repack` subsequently deletes all loose
objects also in packs, these just-loosened promisor objects will be
immediately deleted. However, this extra disk churn is unnecessary in
the first place. For example, in a newly-cloned partial repo that
filters all blob objects (e.g. `--filter=blob:none`), `repack` ends up
unpacking all trees and commits into the filesystem because every
object, in this particular case, is a promisor object. Depending on
the repo size, this increases the disk usage considerably: In my copy
of the linux.git, the object directory peaked 26GB of more disk usage.
In order to avoid this extra disk churn, pass the names of the promisor
packfiles as --keep-pack arguments to the second invocation of
`pack-objects`. This informs `pack-objects` that the promisor objects
are already in a safe packfile and, therefore, do not need to be
loosened.
For testing, we need to validate whether any object was loosened.
However, the "evidence" (loosened objects) is deleted during the
process which prevents us from inspecting the object directory.
Instead, let's teach `pack-objects` to count loosened objects and
emit via trace2 thus allowing inspecting the debug events after the
process is finished. This new event is used on the added regression
test.
Lastly, add a new perf test to evaluate the performance impact
made by this changes (tested on git.git):
Test HEAD^ HEAD
----------------------------------------------------------
5600.3: gc 134.38(41.93+90.95) 7.80(6.72+1.35) -94.2%
For a bigger repository, such as linux.git, the improvement is
even bigger:
Test HEAD^ HEAD
-------------------------------------------------------------------
5600.3: gc 6833.00(918.07+3162.74) 268.79(227.02+39.18) -96.1%
These improvements are particular big because every object in the
newly-cloned partial repository is a promisor object.
Reported-by: SZEDER Gábor <szeder.dev@gmail.com>
Helped-by: Jeff King <peff@peff.net>
Helped-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Rafael Silva <rafaeloliveira.cs@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When --exclude-promisor-objects is given, before traversing any objects
we iterate over all of the objects in any promisor packs, marking them
as UNINTERESTING and SEEN. We turn the oid we get from iterating the
pack into an object with parse_object(), but this has two problems:
- it's slow; we are zlib inflating (and reconstructing from deltas)
every byte of every object in the packfile
- it leaves the tree buffers attached to their structs, which means
our heap usage will grow to store every uncompressed tree
simultaneously. This can be gigabytes.
We can obviously fix the second by freeing the tree buffers after we've
parsed them. But we can observe that the function doesn't look at the
object contents at all! The only reason we call parse_object() is that
we need a "struct object" on which to set the flags. There are two
options here:
- we can look up just the object type via oid_object_info(), and then
call the appropriate lookup_foo() function
- we can call lookup_unknown_object(), which gives us an OBJ_NONE
struct (which will get auto-converted later by object_as_type() via
calls to lookup_commit(), etc).
The first one is closer to the current code, but we do pay the price to
look up the type for each object. The latter should be more efficient in
CPU, though it wastes a little bit of memory (the "unknown" object
structs are a union of all object types, so some of the structs are
bigger than they need to be). It also runs the risk of triggering a
latent bug in code that calls lookup_object() directly but isn't ready
to handle OBJ_NONE (such code would already be buggy, but we use
lookup_unknown_object() infrequently enough that it might be hiding).
I went with the second option here. I don't think the risk is high (and
we'd want to find and fix any such bugs anyway), and it should be more
efficient overall.
The new tests in p5600 show off the improvement (this is on git.git):
Test HEAD^ HEAD
-------------------------------------------------------------------------------
5600.5: count commits 0.37(0.37+0.00) 0.38(0.38+0.00) +2.7%
5600.6: count non-promisor commits 11.74(11.37+0.37) 0.04(0.03+0.00) -99.7%
The improvement is particularly big in this script because _every_
object in the newly-cloned partial repo is a promisor object. So after
marking them all, there's nothing left to traverse.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To get the list of all promisor objects, we not only include all objects
in promisor packs, but also parse each of those objects to see which
objects they reference. After parsing a tree object, the tree->buffer
field will remain populated until we explicitly free it. So in a partial
clone of blob:none, for example, we are essentially reading every tree
in the repository (since they're all in the initial promisor pack), and
keeping all of their uncompressed contents in memory at once.
This patch frees the tree buffers after we've finished marking all of
their reachable objects. We shouldn't need to do this for any other
object type. While we are using some extra memory to store the structs,
no other object type stores the whole contents in its parsed form (we do
sometimes hold on to commit buffers, but less so these days due to
commit graphs, plus most commands which care about promisor objects turn
off the save_commit_buffer global).
Even for a moderate-sized repository like git.git, this patch drops the
peak heap (as measured by massif) for git-fsck from ~1.7GB to ~138MB.
Fsck is a good candidate for measuring here because it doesn't interact
with the promisor code except to call is_promisor_object(), so we can
isolate just this problem.
The added perf test shows only a tiny improvement on my machine for
git.git, since 1.7GB isn't enough to cause any real memory pressure:
Test HEAD^ HEAD
--------------------------------------------------------------------------------
5600.4: fsck 21.26(20.90+0.35) 20.84(20.79+0.04) -2.0%
With linux.git the absolute change is a bit bigger, though still a small
percentage:
Test HEAD^ HEAD
-----------------------------------------------------------------------------
5600.4: fsck 262.26(259.13+3.12) 254.92(254.62+0.29) -2.8%
I didn't have the patience to run it under massif with linux.git, but
it's probably on the order of about 14GB improvement, since that's the
sum of the sizes of all of the uncompressed trees (but still isn't
enough to create memory pressure on this particular machine, which has
64GB of RAM). Smaller machines would probably see a bigger effect on
runtime (and sadly our perf suite does not measure peak heap).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Optimize "rev-list --use-bitmap-index --objects" corner case that
uses negative tags as the stopping points.
* ps/pack-bitmap-optim:
pack-bitmap: avoid traversal of objects referenced by uninteresting tag
p2000-sparse-operations.sh compares different Git commands in
repositories with many files at HEAD but using sparse-checkout to focus
on a small portion of those files.
Add extra copies of the repository that use the sparse-index format so
we can track how that affects the performance of different commands.
At this point in time, the sparse-index is 100% overhead from the CPU
front, and this is measurable in these tests:
Test
---------------------------------------------------------------
2000.2: git status (full-index-v3) 0.59(0.51+0.12)
2000.3: git status (full-index-v4) 0.59(0.52+0.11)
2000.4: git status (sparse-index-v3) 1.40(1.32+0.12)
2000.5: git status (sparse-index-v4) 1.41(1.36+0.08)
2000.6: git add -A (full-index-v3) 2.32(1.97+0.19)
2000.7: git add -A (full-index-v4) 2.17(1.92+0.14)
2000.8: git add -A (sparse-index-v3) 2.31(2.21+0.15)
2000.9: git add -A (sparse-index-v4) 2.30(2.20+0.13)
2000.10: git add . (full-index-v3) 2.39(2.02+0.20)
2000.11: git add . (full-index-v4) 2.20(1.94+0.16)
2000.12: git add . (sparse-index-v3) 2.36(2.27+0.12)
2000.13: git add . (sparse-index-v4) 2.33(2.21+0.16)
2000.14: git commit -a -m A (full-index-v3) 2.47(2.12+0.20)
2000.15: git commit -a -m A (full-index-v4) 2.26(2.00+0.17)
2000.16: git commit -a -m A (sparse-index-v3) 3.01(2.92+0.16)
2000.17: git commit -a -m A (sparse-index-v4) 3.01(2.94+0.15)
Note that there is very little difference between the v3 and v4 index
formats when the sparse-index is enabled. This is primarily due to the
fact that the relative file sizes are the same, and the command time is
mostly taken up by parsing tree objects to expand the sparse index into
a full one.
With the current file layout, the index file sizes are given by this
table:
| full index | sparse index |
+-------------+--------------+
v3 | 108 MiB | 1.6 MiB |
v4 | 80 MiB | 1.2 MiB |
Future updates will improve the performance of Git commands when the
index is sparse.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Create a test script that takes the default performance test (the Git
codebase) and multiplies it by 256 using four layers of duplicated
trees of width four. This results in nearly one million blob entries in
the index. Then, we can clone this repository with sparse-checkout
patterns that demonstrate four copies of the initial repository. Each
clone will use a different index format or mode so peformance can be
tested across the different options.
Note that the initial repo is stripped of submodules before doing the
copies. This preserves the expected data shape of the sparse index,
because directories containing submodules are not collapsed to a sparse
directory entry.
Run a few Git commands on these clones, especially those that use the
index (status, add, commit).
Here are the results on my Linux machine:
Test
--------------------------------------------------------------
2000.2: git status (full-index-v3) 0.37(0.30+0.09)
2000.3: git status (full-index-v4) 0.39(0.32+0.10)
2000.4: git add -A (full-index-v3) 1.42(1.06+0.20)
2000.5: git add -A (full-index-v4) 1.26(0.98+0.16)
2000.6: git add . (full-index-v3) 1.40(1.04+0.18)
2000.7: git add . (full-index-v4) 1.26(0.98+0.17)
2000.8: git commit -a -m A (full-index-v3) 1.42(1.11+0.16)
2000.9: git commit -a -m A (full-index-v4) 1.33(1.08+0.16)
It is perhaps noteworthy that there is an improvement when using index
version 4. This is because the v3 index uses 108 MiB while the v4
index uses 80 MiB. Since the repeated portions of the directories are
very short (f3/f1/f2, for example) this ratio is less pronounced than in
similarly-sized real repositories.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git diff-index" codepath has been taught to trust fsmonitor status
to reduce number of lstat() calls.
* nk/diff-index-fsmonitor:
fsmonitor: add perf test for git diff HEAD
fsmonitor: add assertion that fsmonitor is valid to check_removed
fsmonitor: skip lstat deletion check during git diff-index
"git repack" so far has been only capable of repacking everything
under the sun into a single pack (or split by size). A cleverer
strategy to reduce the cost of repacking a repository has been
introduced.
* tb/geometric-repack:
builtin/pack-objects.c: ignore missing links with --stdin-packs
builtin/repack.c: reword comment around pack-objects flags
builtin/repack.c: be more conservative with unsigned overflows
builtin/repack.c: assign pack split later
t7703: test --geometric repack with loose objects
builtin/repack.c: do not repack single packs with --geometric
builtin/repack.c: add '--geometric' option
packfile: add kept-pack cache for find_kept_pack_entry()
builtin/pack-objects.c: rewrite honor-pack-keep logic
p5303: measure time to repack with keep
p5303: add missing &&-chains
builtin/pack-objects.c: add '--stdin-packs' option
revision: learn '--no-kept-objects'
packfile: introduce 'find_kept_pack_entry()'
Perf test update to work better in secondary worktrees.
* jk/perf-in-worktrees:
t/perf: avoid copying worktree files from test repo
t/perf: handle worktrees as test repos
When preparing the bitmap walk, we first establish the set of of have
and want objects by iterating over the set of pending objects: if an
object is marked as uninteresting, it's declared as an object we already
have, otherwise as an object we want. These two sets are then used to
compute which transitively referenced objects we need to obtain.
One special case here are tag objects: when a tag is requested, we
resolve it to its first not-tag object and add both resolved objects as
well as the tag itself into either the have or want set. Given that the
uninteresting-property always propagates to referenced objects, it is
clear that if the tag is uninteresting, so are its children and vice
versa. But we fail to propagate the flag, which effectively means that
referenced objects will always be interesting except for the case where
they have already been marked as uninteresting explicitly.
This mislabeling does not impact correctness: we now have it in our
"wants" set, and given that we later do an `AND NOT` of the bitmaps of
"wants" and "haves" sets it is clear that the result must be the same.
But we now start to needlessly traverse the tag's referenced objects in
case it is uninteresting, even though we know that each referenced
object will be uninteresting anyway. In the worst case, this can lead to
a complete graph walk just to establish that we do not care for any
object.
Fix the issue by propagating the `UNINTERESTING` flag to pointees of tag
objects and add a benchmark with negative revisions to p5310. This shows
some nice performance benefits, tested with linux.git:
Test HEAD~ HEAD
---------------------------------------------------------------------------------------------------------------
5310.3: repack to disk 193.18(181.46+16.42) 194.61(183.41+15.83) +0.7%
5310.4: simulated clone 25.93(24.88+1.05) 25.81(24.73+1.08) -0.5%
5310.5: simulated fetch 2.64(5.30+0.69) 2.59(5.16+0.65) -1.9%
5310.6: pack to file (bitmap) 58.75(57.56+6.30) 58.29(57.61+5.73) -0.8%
5310.7: rev-list (commits) 1.45(1.18+0.26) 1.46(1.22+0.24) +0.7%
5310.8: rev-list (objects) 15.35(14.22+1.13) 15.30(14.23+1.07) -0.3%
5310.9: rev-list with tag negated via --not --all (objects) 22.49(20.93+1.56) 0.11(0.09+0.01) -99.5%
5310.10: rev-list with negative tag (objects) 0.61(0.44+0.16) 0.51(0.35+0.16) -16.4%
5310.11: rev-list count with blob:none 12.15(11.19+0.96) 12.18(11.19+0.99) +0.2%
5310.12: rev-list count with blob:limit=1k 17.77(15.71+2.06) 17.75(15.63+2.12) -0.1%
5310.13: rev-list count with tree:0 1.69(1.31+0.38) 1.68(1.28+0.39) -0.6%
5310.14: simulated partial clone 20.14(19.15+0.98) 19.98(18.93+1.05) -0.8%
5310.16: clone (partial bitmap) 12.78(13.89+1.07) 12.72(13.99+1.01) -0.5%
5310.17: pack to file (partial bitmap) 42.07(45.44+2.72) 41.44(44.66+2.80) -1.5%
5310.18: rev-list with tree filter (partial bitmap) 0.44(0.29+0.15) 0.46(0.32+0.14) +4.5%
While most benchmarks are probably in the range of noise, the newly
added 5310.9 and 5310.10 benchmarks consistenly perform better.
Signed-off-by: Patrick Steinhardt <ps@pks.im>.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Preliminary changes to fsmonitor integration.
* jh/fsmonitor-prework:
fsmonitor: refactor initialization of fsmonitor_last_update token
fsmonitor: allow all entries for a folder to be invalidated
fsmonitor: log FSMN token when reading and writing the index
fsmonitor: log invocation of FSMonitor hook to trace2
read-cache: log the number of scanned files to trace2
read-cache: log the number of lstat calls to trace2
preload-index: log the number of lstat calls to trace2
p7519: add trace logging during perf test
p7519: move watchman cleanup earlier in the test
p7519: fix watchman watch-list test on Windows
p7519: do not rely on "xargs -d" in test
When running the perf suite, we copy files from an existing $GIT_DIR to
a scratch repository to give us a realistic setup on which to operate.
Since the perf scripts themselves may modify the scratch repository, we
want to make sure we've scrubbed any references back to the original.
One existing example is that we avoid copying the file "commondir" at
the top-level of the repository. In a worktree git-dir (e.g.,
.git/worktrees/foo), that file contains the path to the parent
repository; copying it could mean ref updates in the scratch repository
affect the original.
But there are other files we should cover, too:
- "gitdir" in a worktree git-dir contains the path to the actual .git
file in the working tree. We _shouldn't_ end up looking at it at
all, since the lack of a "commondir" file means Git won't consider
this to be a worktree git-dir. But it's best to err on the safe
side.
- in a parent repository that contains worktrees, the
"$GIT_DIR/worktrees" directory will contain the git dirs for the
individual worktrees. Which will themselves contain commondir and
gitdir files that may reference the original repository. We should
likewise remove them.
Note that this does mean that the perf suite's scratch repositories
will never have any worktrees. That's OK; we don't have any perf tests
that are influenced by their presence. If we add any, they'd
probably want to create the worktrees themselves anyway.
This patch adds both paths to the set of omissions in
test_perf_copy_repo_contents(). Note that we won't get confused here by
matching arbitrary names like refs/heads/commondir. This list is always
matching top-level entries in $GIT_DIR (we rely on "cp -R" to do the
actual recursion).
Suggested-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: Jeff King <peff@peff.net>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The perf suite gets confused when test_perf_default_repo is pointed at a
worktree (which includes when it is run from within a worktree at all,
since the default is to use the current repository).
Here's an example:
$ git worktree add ~/foo
Preparing worktree (new branch 'foo')
HEAD is now at 328c109303 The eighth batch
$ cd ~/foo
$ make
[...build output...]
$ cd t/perf
$ ./p0000-perf-lib-sanity.sh -v -i
[...]
perf 1 - test_perf_default_repo works:
running:
foo=$(git rev-parse HEAD) &&
test_export foo
fatal: ambiguous argument 'HEAD': unknown revision or path not in the working tree.
Use '--' to separate paths from revisions, like this:
'git <command> [<revision>...] -- [<file>...]'
The problem is that we didn't copy all of the necessary files from the
source repository (in this case we got HEAD, but we have no refs!). We
discover the git-dir with "rev-parse --git-dir", but this points to the
worktree's partial repository in .../.git/worktrees/foo.
That partial repository has a "commondir" file which points to the main
repository, where the actual refs are stored, but we don't copy it. This
is the correct thing to do, though! If we did copy it, then our scratch
test repo would be pointing back to the original main repo, and any ref
updates we made in the tests would impact that original repo.
Instead, we need to either:
1. Make a scratch copy of the original main repo (in addition to the
worktree repo), and point the scratch worktree repo's commondir at
it. This preserves the original relationship, but it's doubtful any
script really cares (if they are testing worktree performance,
they'd probably make their own worktrees). And it's trickier to get
right.
2. Collapse the main and worktree repos into a single scratch repo.
This can be done by copying everything from both, preferring any
files from the worktree repo.
This patch does the second one. With this applied, the example above
results in p0000 running successfully.
Reported-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Jeff King <peff@peff.net>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add two new tests to measure repack performance. Both tests split the
repository into synthetic "pushes", and then leave the remaining objects
in a big base pack.
The first new test marks an empty pack as "kept" and then passes
--honor-pack-keep to avoid including objects in it. That doesn't change
the resulting pack, but it does let us compare to the normal repack case
to see how much overhead we add to check whether objects are kept or
not.
The other test is of --stdin-packs, which gives us a sense of how that
number scales based on the number of packs we provide as input. In each
of those tests, the empty pack isn't considered, but the residual pack
(objects that were left over and not included in one of the synthetic
push packs) is marked as kept.
(Note that in the single-pack case of the --stdin-packs test, there is
nothing do since there are no non-excluded packs).
Here are some timings on a recent clone of the kernel:
5303.5: repack (1) 57.26(54.59+10.84)
5303.6: repack with kept (1) 57.33(54.80+10.51)
in the 50-pack case, things start to slow down:
5303.11: repack (50) 71.54(88.57+4.84)
5303.12: repack with kept (50) 85.12(102.05+4.94)
and by the time we hit 1,000 packs, things are substantially worse, even
though the resulting pack produced is the same:
5303.17: repack (1000) 216.87(490.79+14.57)
5303.18: repack with kept (1000) 665.63(938.87+15.76)
That's because the code paths around handling .keep files are known to
scale badly; they look in every single pack file to find each object.
Our solution to that was to notice that most repos don't have keep
files, and to make that case a fast path. But as soon as you add a
single .keep, that part of pack-objects slows down again (even if we
have fewer objects total to look at).
Likewise, the scaling is pretty extreme on --stdin-packs (but each
subsequent test is also being asked to do more work):
5303.7: repack with --stdin-packs (1) 0.01(0.01+0.00)
5303.13: repack with --stdin-packs (50) 3.53(12.07+0.24)
5303.19: repack with --stdin-packs (1000) 195.83(371.82+8.10)
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These are in a helper function, so the usual chain-lint doesn't notice
them. This function is still not perfect, as it has some git invocations
on the left-hand-side of the pipe, but it's primary purpose is timing,
not finding bugs or correctness issues.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add optional trace logging to allow us to better compare performance of
various fsmonitor providers and compare results with non-fsmonitor runs.
Currently, this includes Trace2 logging, but may be extended to include
other trace targets, such as GIT_TRACE_FSMONITOR if desired.
Using this logging helped me explain an odd behavior on MacOS where the
kernel was dropping events and causing the hook to Watchman to timeout.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Shutdown Watchman after the Watchman-based tests and before the block of
"no fsmonitor" tests.
This helps ensure that Watchman cannot affect the test results for the
other.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Only use the final portion of the test trash directory file name
when verifying that Watchman was started.
On Windows and under the SDK, $GIT_WORKTREE is a cygwin-style
path with forward slashes and a "/c/" drive name. However
`watchman watch-list` reports a proper Windows-style pathname
with drive letters and backslashes. This causes the grep to
fail. Since we don't really care about the full pathname (and
we really don't want to bother with normalizaing them), just see
if the test-name portion of the path is found.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert the test to use a more portable method to update the mtime on a
large number of files under version control.
The Mac version of xargs does not support the "-d" option.
Likewise, the "-0" and "--null" options are not portable.
Furthermore, use `test-tool chmtime` rather than `touch` to update the
mtime to ensure that it is actually updated (especially on file systems
with only whole second resolution).
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some pretty-format specifiers do not need the data in commit object
(e.g. "%H"), but we were over-eager to load and parse it, which has
been made even lazier.
* jk/pretty-lazy-load-commit:
pretty: lazy-load commit data when expanding user-format
Using "1~5" isn't portable. Nobody seems to have noticed, since perhaps
people don't tend to run the perf suite on more exotic platforms. Still,
it's better to set a good example.
We can use:
perl -ne 'print if $. % 5 == 1'
instead. But we can further observe that perl does a good job of the
other parts of this pipeline, and fold the whole thing together.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we expand a user-format, we try to avoid work that isn't necessary
for the output. For instance, we don't bother parsing the commit header
until we know we need the author, subject, etc.
But we do always load the commit object's contents from disk, even if
the format doesn't require it (e.g., just "%H"). Traditionally this
didn't matter much, because we'd have loaded it as part of the traversal
anyway, and we'd typically have those bytes attached to the commit
struct (or these days, cached in a commit-slab).
But when we have a commit-graph, we might easily get to the point of
pretty-printing a commit without ever having looked at the actual object
contents. We should push off that load (and reencoding) until we're
certain that it's needed.
I think the results of p4205 show the advantage pretty clearly (we serve
parent and tree oids out of the commit struct itself, so they benefit as
well):
# using git.git as the test repo
Test HEAD^ HEAD
----------------------------------------------------------------------
4205.1: log with %H 0.40(0.39+0.01) 0.03(0.02+0.01) -92.5%
4205.2: log with %h 0.45(0.44+0.01) 0.09(0.09+0.00) -80.0%
4205.3: log with %T 0.40(0.39+0.00) 0.04(0.04+0.00) -90.0%
4205.4: log with %t 0.46(0.46+0.00) 0.09(0.08+0.01) -80.4%
4205.5: log with %P 0.39(0.39+0.00) 0.03(0.03+0.00) -92.3%
4205.6: log with %p 0.46(0.46+0.00) 0.10(0.09+0.00) -78.3%
4205.7: log with %h-%h-%h 0.52(0.51+0.01) 0.15(0.14+0.00) -71.2%
4205.8: log with %an-%ae-%s 0.42(0.41+0.00) 0.42(0.41+0.01) +0.0%
# using linux.git as the test repo
Test HEAD^ HEAD
----------------------------------------------------------------------
4205.1: log with %H 7.12(6.97+0.14) 0.76(0.65+0.11) -89.3%
4205.2: log with %h 7.35(7.19+0.16) 1.30(1.19+0.11) -82.3%
4205.3: log with %T 7.58(7.42+0.15) 1.02(0.94+0.08) -86.5%
4205.4: log with %t 8.05(7.89+0.15) 1.55(1.41+0.13) -80.7%
4205.5: log with %P 7.12(7.01+0.10) 0.76(0.69+0.07) -89.3%
4205.6: log with %p 7.38(7.27+0.10) 1.32(1.20+0.12) -82.1%
4205.7: log with %h-%h-%h 7.81(7.67+0.13) 1.79(1.67+0.12) -77.1%
4205.8: log with %an-%ae-%s 7.90(7.74+0.15) 7.81(7.66+0.15) -1.1%
I added the final test to show where we don't improve (the 1% there is
just lucky noise), but also as a regression test to make sure we're not
doing anything stupid like loading the commit multiple times when there
are several placeholders that need it.
Reported-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
p7519 measures the performance of the fsmonitor code. To do this, it
uses the installed copy of Watchman. If Watchman isn't installed, a noop
integration script is installed in its place.
When in the latter mode, it is expected that the script should not write
a "last update token": in fact, it doesn't write anything at all since
the script is blank.
Commit 33226af42b (t/perf/fsmonitor: improve error message if typoing
hook name, 2020-10-26) made sure that running 'git update-index
--fsmonitor' did not write anything to stderr, but this is not the case
when using the empty Watchman script, since Git will complain that:
$ which watchman
watchman not found
$ cat .git/hooks/fsmonitor-empty
$ git -c core.fsmonitor=.git/hooks/fsmonitor-empty update-index --fsmonitor
warning: Empty last update token.
Prior to 33226af42b, the output wasn't checked at all, which allowed
this noop mode to work. But, 33226af42b breaks p7519 when running it
without a 'watchman(1)' on your system.
Handle this by only checking that the stderr is empty only when running
with a real watchman executable. Otherwise, assert that the error
message is the expected one when running in the noop mode.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Acked-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
test_export() has been self-recursive since its inception even though a
simple for-loop would have served just as well to append its arguments
to the `test_export_` variable separated by the pipe character "|".
Recently `test_export_` was changed instead to a space-separated list of
tokens to be exported, an operation which can be accomplished via a
single simple assignment, with no need for looping or recursion.
Therefore, simplify the implementation.
While at it, take advantage of the fact that variable names to be
exported are shell identifiers, thus won't be composed of special
characters or whitespace, thus simple a `$*` can be used rather than
magical `"$@"`.
Signed-off-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
test_perf() runs each test in its own subshell which makes it difficult
to persist variables between tests. test_export() addresses this
shortcoming by grabbing the values of specified variables after a test
runs but before the subshell exits, and writes those values to a file
which is loaded into the environment of subsequent tests.
To grab the values to be persisted, test_export() pipes the output of
the shell's builtin `set` command through `sed` which plucks them out
using a regular expression along the lines of `s/^(var1|var2)/.../p`.
Unfortunately, though, this use of alternation is not portable. For
instance, BSD-lineage `sed` (including macOS `sed`) does not support it
in the default "basic regular expression" mode (BRE). It may be possible
to enable "extended regular expression" mode (ERE) in some cases with
`sed -E`, however, `-E` is neither portable nor part of POSIX.
Fortunately, alternation is unnecessary in this case and can easily be
avoided, so replace it with a series of simple expressions such as
`s/^var1/.../p;s/^var2/.../p`.
While at it, tighten the expressions so they match the variable names
exactly rather than matching prefixes (i.e. use `s/^var1=/.../p`).
If the requirements of test_export() become more complex in the future,
then an alternative would be to replace `sed` with `perl` which supports
alternation on all platforms, however, the simple elimination of
alternation via multiple `sed` expressions suffices for the present.
Reported-by: Sangeeta <sangunb09@gmail.com>
Diagnosed-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git update-ref --stdin" learns to take multiple transactions in a
single session.
* ps/update-ref-multi-transaction:
update-ref: disallow "start" for ongoing transactions
p1400: use `git-update-ref --stdin` to test multiple transactions
update-ref: allow creation of multiple transactions
t1400: avoid touching refs on filesystem
Simplify test and make error messages more clear here.
Per feedback from Junio in
33226af42b (t/perf/fsmonitor: improve error message if typoing hook
name, 2020-10-26)
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Acked-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In commit 0a0fbbe3ff (refs: remove lookup cache for
reference-transaction hook, 2020-08-25), a new benchmark was added to
p1400 which has the intention to exercise creation of multiple
transactions in a single process. As git-update-ref wasn't yet able to
create multiple transactions with a single run we instead used git-push.
As its non-atomic version creates a transaction per reference update,
this was the best approximation we could make at that point in time.
Now that `git-update-ref --stdin` supports creation of multiple
transactions, let's convert the benchmark to use that instead. It has
less overhead and it's also a lot clearer what the actual intention is.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This benchmark covers the git status time for a heavily
dirty directory - benchmarking fsmonitor's refresh
When running to compare our perl vs rs-git-fsmonitor - we see that
the perl script incurs significant overhead - further motivation
to provide a faster implementation within git.
7519.7: status (dirty) (fsmonitor=query-watchman) 10.05(7.78+1.56)
7519.20: status (dirty) (fsmonitor=rs-git-fsmonitor) 6.72(4.37+1.64)
7519.33: status (dirty) (fsmonitor=disabled) 5.62(4.24+2.03)
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This prepares for it being called multiple times when
testing different hooks
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is extremely verbose, printing >10K non-useful lines
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The full name is lengthy and makes it hard to read
Before:
7519.3: status (fsmonitor=/home/nipunn/src/server/.git/hooks/rs-git-fsmonitor) 0.02(0.01+0.00)
After
7519.3: status (fsmonitor=rs-git-fsmonitor) 0.03(0.02+0.00)
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There was much duplication here. Prepares for making
changes to the description.
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Previously - it would silently run the perf suite w/o using
fsmonitor - fsmonitor errors are not hard failures.
Now it errors loudly.
GIT_PERF_7519_FSMONITOR="$HOME/rs-git-fsmonitorr"
./p7519-fsmonitor.sh -i -v
fatal: cannot run /home/nipunn/rs-git-fsmonitorr:
No such file or directory
not ok 2 - setup for fsmonitor
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is only required to be set up once. This prepares for
testing multiple hooks in one invocation.
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In preparation for testing multiple fsmonitor hooks
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Much of the benchmark code is redundant. This is
easier to understand and edit.
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Perf tests have not been linted for some time.
They've grown some seq instead of test_seq. This
runs the existing lints on the perf tests as well.
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Results for the git-diff fsmonitor optimization
in patch in the parent-rev (using a 400k file repo to test)
As you can see here - git diff with fsmonitor running is
significantly better with this patch series (80% faster on my
workload)!
GIT_PERF_LARGE_REPO=~/src/server ./run v2.29.0-rc1 . -- p7519-fsmonitor.sh
Test v2.29.0-rc1 this tree
-----------------------------------------------------------------------------------------------------------------
7519.2: status (fsmonitor=.git/hooks/fsmonitor-watchman) 1.46(0.82+0.64) 1.47(0.83+0.62) +0.7%
7519.3: status -uno (fsmonitor=.git/hooks/fsmonitor-watchman) 0.16(0.12+0.04) 0.17(0.12+0.05) +6.3%
7519.4: status -uall (fsmonitor=.git/hooks/fsmonitor-watchman) 1.36(0.73+0.62) 1.37(0.76+0.60) +0.7%
7519.5: diff (fsmonitor=.git/hooks/fsmonitor-watchman) 0.85(0.22+0.63) 0.14(0.10+0.05) -83.5%
7519.6: diff -- 0_files (fsmonitor=.git/hooks/fsmonitor-watchman) 0.12(0.08+0.05) 0.13(0.11+0.02) +8.3%
7519.7: diff -- 10_files (fsmonitor=.git/hooks/fsmonitor-watchman) 0.12(0.08+0.04) 0.13(0.09+0.04) +8.3%
7519.8: diff -- 100_files (fsmonitor=.git/hooks/fsmonitor-watchman) 0.12(0.07+0.05) 0.13(0.07+0.06) +8.3%
7519.9: diff -- 1000_files (fsmonitor=.git/hooks/fsmonitor-watchman) 0.12(0.09+0.04) 0.13(0.08+0.05) +8.3%
7519.10: diff -- 10000_files (fsmonitor=.git/hooks/fsmonitor-watchman) 0.14(0.09+0.05) 0.13(0.10+0.03) -7.1%
7519.12: status (fsmonitor=) 1.67(0.93+1.49) 1.67(0.99+1.42) +0.0%
7519.13: status -uno (fsmonitor=) 0.37(0.30+0.82) 0.37(0.33+0.79) +0.0%
7519.14: status -uall (fsmonitor=) 1.58(0.97+1.35) 1.57(0.86+1.45) -0.6%
7519.15: diff (fsmonitor=) 0.34(0.28+0.83) 0.34(0.27+0.83) +0.0%
7519.16: diff -- 0_files (fsmonitor=) 0.09(0.06+0.04) 0.09(0.08+0.02) +0.0%
7519.17: diff -- 10_files (fsmonitor=) 0.09(0.07+0.03) 0.09(0.06+0.05) +0.0%
7519.18: diff -- 100_files (fsmonitor=) 0.09(0.06+0.04) 0.09(0.06+0.04) +0.0%
7519.19: diff -- 1000_files (fsmonitor=) 0.09(0.06+0.04) 0.09(0.05+0.05) +0.0%
7519.20: diff -- 10000_files (fsmonitor=) 0.10(0.08+0.04) 0.10(0.06+0.05) +0.0%
I also added a benchmark for a tiny git diff workload w/ a pathspec.
I see an approximately .02 second overhead added w/ and w/o fsmonitor
From looking at these results, I suspected that refresh_fsmonitor
is already happening during git diff - independent of this patch
series' optimization. Confirmed that suspicion by breaking on
refresh_fsmonitor.
(gdb) bt [simplified]
0 refresh_fsmonitor at fsmonitor.c:176
1 ie_match_stat at read-cache.c:375
2 match_stat_with_submodule at diff-lib.c:237
4 builtin_diff_files at builtin/diff.c:260
5 cmd_diff at builtin/diff.c:541
6 run_builtin at git.c:450
7 handle_builtin at git.c:700
8 run_argv at git.c:767
9 cmd_main at git.c:898
10 main at common-main.c:52
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The first git status would be inflated due to warming of
filesystem cache. This makes the results comparable.
Before
Test this tree
--------------------------------------------------------------------------------
7519.2: status (fsmonitor=.git/hooks/fsmonitor-watchman) 2.52(1.59+1.56)
7519.3: status -uno (fsmonitor=.git/hooks/fsmonitor-watchman) 0.18(0.12+0.06)
7519.4: status -uall (fsmonitor=.git/hooks/fsmonitor-watchman) 1.36(0.73+0.62)
7519.7: status (fsmonitor=) 0.69(0.52+0.90)
7519.8: status -uno (fsmonitor=) 0.37(0.28+0.81)
7519.9: status -uall (fsmonitor=) 1.53(0.93+1.32)
After
Test this tree
--------------------------------------------------------------------------------
7519.2: status (fsmonitor=.git/hooks/fsmonitor-watchman) 0.39(0.33+0.06)
7519.3: status -uno (fsmonitor=.git/hooks/fsmonitor-watchman) 0.17(0.13+0.05)
7519.4: status -uall (fsmonitor=.git/hooks/fsmonitor-watchman) 1.34(0.77+0.56)
7519.7: status (fsmonitor=) 0.70(0.53+0.90)
7519.8: status -uno (fsmonitor=) 0.37(0.32+0.78)
7519.9: status -uall (fsmonitor=) 1.55(1.01+1.25)
Signed-off-by: Nipunn Koorapati <nipunn@dropbox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There is a logic to estimate how many objects are in the
repository, which is mean to run once per process invocation, but
it ran every time the estimated value was requested.
* jk/dont-count-existing-objects-twice:
packfile: actually set approximate_object_count_valid
The approximate_object_count() function tries to compute the count only
once per process. But ever since it was introduced in 8e3f52d778
(find_unique_abbrev: move logic out of get_short_sha1(), 2016-10-03), we
failed to actually set the "valid" flag, meaning we'd compute it fresh
on every call.
This turns out not to be _too_ bad, because we're only iterating through
the packed_git list, and not making any system calls. But since it may
get called for every abbreviated hash we output, even this can add up if
you have many packs.
Here are before-and-after timings for a new perf test which just asks
rev-list to abbreviate each commit hash (the test repo is linux.git,
with commit-graphs):
Test origin HEAD
----------------------------------------------------------------------------
5303.3: rev-list (1) 28.91(28.46+0.44) 29.03(28.65+0.38) +0.4%
5303.4: abbrev-commit (1) 1.18(1.06+0.11) 1.17(1.02+0.14) -0.8%
5303.7: rev-list (50) 28.95(28.56+0.38) 29.50(29.17+0.32) +1.9%
5303.8: abbrev-commit (50) 3.67(3.56+0.10) 3.57(3.42+0.15) -2.7%
5303.11: rev-list (1000) 30.34(29.89+0.43) 30.82(30.35+0.46) +1.6%
5303.12: abbrev-commit (1000) 86.82(86.52+0.29) 77.82(77.59+0.22) -10.4%
5303.15: load 10,000 packs 0.08(0.02+0.05) 0.08(0.02+0.06) +0.0%
It doesn't help at all when we have 1 pack (5303.4), but we get a 10%
speedup when there are 1000 packs (5303.12). That's a modest speedup for
a case that's already slow and we'd hope to avoid in general (note how
slow it is even after, because we have to look in each of those packs
for abbreviations). But it's a one-line change that clearly matches the
original intent, so it seems worth doing.
The included perf test may also be useful for keeping an eye on any
regressions in the overall abbreviation code.
Reported-by: Rasmus Villemoes <rv@rasmusvillemoes.dk>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When adding the reference-transaction hook, there were concerns about
the performance impact it may have on setups which do not make use of
the new hook at all. After all, it gets executed every time a reftx is
prepared, committed or aborted, which linearly scales with the number of
reference-transactions created per session. And as there are code paths
like `git push` which create a new transaction for each reference to be
updated, this may translate to calling `find_hook()` quite a lot.
To address this concern, a cache was added with the intention to not
repeatedly do negative hook lookups. Turns out this cache caused a
regression, which was fixed via e5256c82e5 (refs: fix interleaving hook
calls with reference-transaction hook, 2020-08-07). In the process of
discussing the fix, we realized that the cache doesn't really help even
in the negative-lookup case. While performance tests added to benchmark
this did show a slight improvement in the 1% range, this really doesn't
warrent having a cache. Furthermore, it's quite flaky, too. E.g. running
it twice in succession produces the following results:
Test master pks-reftx-hook-remove-cache
--------------------------------------------------------------------------
1400.2: update-ref 2.79(2.16+0.74) 2.73(2.12+0.71) -2.2%
1400.3: update-ref --stdin 0.22(0.08+0.14) 0.21(0.08+0.12) -4.5%
Test master pks-reftx-hook-remove-cache
--------------------------------------------------------------------------
1400.2: update-ref 2.70(2.09+0.72) 2.74(2.13+0.71) +1.5%
1400.3: update-ref --stdin 0.21(0.10+0.10) 0.21(0.08+0.13) +0.0%
One case notably absent from those benchmarks is a single executable
searching for the hook hundreds of times, which is exactly the case for
which the negative cache was added. p1400.2 will spawn a new update-ref
for each transaction and p1400.3 only has a single reference-transaction
for all reference updates. So this commit adds a third benchmark, which
performs an non-atomic push of a thousand references. This will create a
new reference transaction per reference. But even for this case, the
negative cache doesn't consistently improve performance:
Test master pks-reftx-hook-remove-cache
--------------------------------------------------------------------------
1400.4: nonatomic push 6.63(6.50+0.13) 6.81(6.67+0.14) +2.7%
1400.4: nonatomic push 6.35(6.21+0.14) 6.39(6.23+0.16) +0.6%
1400.4: nonatomic push 6.43(6.31+0.13) 6.42(6.28+0.15) -0.2%
So let's just remove the cache altogether to simplify the code.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When PERF_EXTRA is enabled, p5302 checks the performance of index-pack
with various numbers of threads. This can be useful for deciding what
the default should be (which is currently capped at 3 threads based on
the results of this script).
However, we only go up to 8 threads, and modern machines may have more.
Let's get the number of CPUs from test-tool, and test various numbers of
threads between one and that maximum.
Note that the current tests aren't all identical, as we have to set
GIT_FORCE_THREADS for the --threads=1 test (which measures the overhead
of starting a single worker thread versus the "0" case of using the main
thread). To keep the loop simple, we'll keep the "0" case out of it, and
set GIT_FORCE_THREADS=1 for all of the other cases (it's a noop for all
but the "1" case, since numbers higher than 1 would always need
threads).
Note also that we could skip running "test-tool" if PERF_EXTRA isn't
set. However, there's some small value in knowing the number of threads,
so that we can mark each test as skipped in the output.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The primary function of the perf suite is to detect regressions (or
improvements) between versions of Git. The only numbers we show a direct
comparison for are timings between the same test run on two different
versions.
However, it can sometimes be used to collect other information. For
instance, p5302 runs the same index-pack operation with different thread
counts. The output doesn't directly compare these, but anybody
interested in working on index-pack can manually compare the results.
For a normal regression run of the full perf-suite, though, this incurs
a significant cost to generate numbers nobody will actually look at;
about 25% of the total time of the test suite is spent in p5302. And the
low-thread-count runs are the most expensive part of it, since they're
(unsurprisingly) not using as many threads.
Let's skip these tests by default, but make it possible for people
working on index-pack to still run them by setting an environment
variable. Rather than make this specific to p5302, let's introduce a
generic mechanism. This makes it possible to run the full suite with
every possible test if somebody really wants to burn some CPU.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The low-level reference transactions used to update references are
currently completely opaque to the user. While certainly desirable in
most usecases, there are some which might want to hook into the
transaction to observe all queued reference updates as well as observing
the abortion or commit of a prepared transaction.
One such usecase would be to have a set of replicas of a given Git
repository, where we perform Git operations on all of the repositories
at once and expect the outcome to be the same in all of them. While
there exist hooks already for a certain subset of Git commands that
could be used to implement a voting mechanism for this, many others
currently don't have any mechanism for this.
The above scenario is the motivation for the new "reference-transaction"
hook that reaches directly into Git's reference transaction mechanism.
The hook receives as parameter the current state the transaction was
moved to ("prepared", "committed" or "aborted") and gets via its
standard input all queued reference updates. While the exit code gets
ignored in the "committed" and "aborted" states, a non-zero exit code in
the "prepared" state will cause the transaction to be aborted
prematurely.
Given the usecase described above, a voting mechanism can now be
implemented via this hook: as soon as it gets called, it will take all
of stdin and use it to cast a vote to a central service. When all
replicas of the repository agree, the hook will exit with zero,
otherwise it will abort the transaction by returning non-zero. The most
important upside is that this will catch _all_ commands writing
references at once, allowing to implement strong consistency for
reference updates via a single mechanism.
In order to test the impact on the case where we don't have any
"reference-transaction" hook installed in the repository, this commit
introduce two new performance tests for git-update-refs(1). Run against
an empty repository, it produces the following results:
Test origin/master HEAD
--------------------------------------------------------------------
1400.2: update-ref 2.70(2.10+0.71) 2.71(2.10+0.73) +0.4%
1400.3: update-ref --stdin 0.21(0.09+0.11) 0.21(0.07+0.14) +0.0%
The performance test p1400.2 creates, updates and deletes a branch a
thousand times, thus averaging runtime of git-update-refs over 3000
invocations. p1400.3 instead calls `git-update-refs --stdin` three times
and queues a thousand creations, updates and deletes respectively.
As expected, p1400.3 consistently shows no noticeable impact, as for
each batch of updates there's a single call to access(3P) for the
negative hook lookup. On the other hand, for p1400.2, one can see an
impact caused by this patchset. But doing five runs of the performance
tests where each one was run with GIT_PERF_REPEAT_COUNT=10, the overhead
ranged from -1.5% to +1.1%. These inconsistent performance numbers can
be explained by the overhead of spawning 3000 processes. This shows that
the overhead of assembling the hook path and executing access(3P) once
to check if it's there is mostly outweighed by the operating system's
overhead.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Sometimes a bitmap traversal still has to walk some commits manually,
because those commits aren't included in the bitmap packfile (e.g., due
to a push or commit since the last full repack). If we're given an
object filter, we don't pass it down to this traversal. It's not
necessary for correctness because the bitmap code has its own filters to
post-process the bitmap result (which it must, to filter out the objects
that _are_ mentioned in the bitmapped packfile).
And with blob filters, there was no performance reason to pass along
those filters, either. The fill-in traversal could omit them from the
result, but it wouldn't save us any time to do so, since we'd still have
to walk each tree entry to see if it's a blob or not.
But now that we support tree filters, there's opportunity for savings. A
tree:depth=0 filter means we can avoid accessing trees entirely, since
we know we won't them (or any of the subtrees or blobs they point to).
The new test in p5310 shows this off (the "partial bitmap" state is one
where HEAD~100 and its ancestors are all in a bitmapped pack, but
HEAD~100..HEAD are not). Here are the results (run against linux.git):
Test HEAD^ HEAD
-------------------------------------------------------------------------------------------------
[...]
5310.16: rev-list with tree filter (partial bitmap) 0.19(0.17+0.02) 0.03(0.02+0.01) -84.2%
The absolute number of savings isn't _huge_, but keep in mind that we
only omitted 100 first-parent links (in the version of linux.git here,
that's 894 actual commits). In a more pathological case, we might have a
much larger proportion of non-bitmapped commits. I didn't bother
creating such a case in the perf script because the setup is expensive,
and this is plenty to show the savings as a percentage.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the previous patch, we made it easy to define other filters that
exclude all objects of a certain type. Use that in order to implement
bitmap-level filtering for the '--filter=tree:<n>' filter when 'n' is
equal to 0.
The general case is not helped by bitmaps, since for values of 'n > 0',
the object filtering machinery requires a full-blown tree traversal in
order to determine the depth of a given tree. Caching this is
non-obvious, too, since the same tree object can have a different depth
depending on the context (e.g., a tree was moved up in the directory
hierarchy between two commits).
But, the 'n = 0' case can be helped, and this patch does so. Running
p5310.11 in this tree and on master with the kernel, we can see that
this case is helped substantially:
Test master this tree
--------------------------------------------------------------------------------
5310.11: rev-list count with tree:0 10.68(10.39+0.27) 0.06(0.04+0.01) -99.4%
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The custom hash function used by "git fast-import" has been
replaced with the one from hashmap.c, which gave us a nice
performance boost.
* jk/fast-import-use-hashmap:
fast-import: replace custom hash with hashmap.c
We use a custom hash in fast-import to store the set of objects we've
imported so far. It has a fixed set of 2^16 buckets and chains any
collisions with a linked list. As the number of objects grows larger
than that, the load factor increases and we degrade to O(n) lookups and
O(n^2) insertions.
We can scale better by using our hashmap.c implementation, which will
resize the bucket count as we grow. This does incur an extra memory cost
of 8 bytes per object, as hashmap stores the integer hash value for each
entry in its hashmap_entry struct (which we really don't care about
here, because we're just reusing the embedded object hash). But I think
the numbers below justify this (and our per-object memory cost is
already much higher).
I also looked at using khash, but it seemed to perform slightly worse
than hashmap at all sizes, and worse even than the existing code for
small sizes. It's also awkward to use here, because we want to look up a
"struct object_entry" from a "struct object_id", and it doesn't handle
mismatched keys as well. Making a mapping of object_id to object_entry
would be more natural, but that would require pulling the embedded oid
out of the object_entry or incurring an extra 32 bytes per object.
In a synthetic test creating as many cheap, tiny objects as possible
perl -e '
my $bits = shift;
my $nr = 2**$bits;
for (my $i = 0; $i < $nr; $i++) {
print "blob\n";
print "data 4\n";
print pack("N", $i);
}
' $bits | git fast-import
I got these results:
nr_objects master khash hashmap
2^20 0m4.317s 0m5.109s 0m3.890s
2^21 0m10.204s 0m9.702s 0m7.933s
2^22 0m27.159s 0m17.911s 0m16.751s
2^23 1m19.038s 0m35.080s 0m31.963s
2^24 4m18.766s 1m10.233s 1m6.793s
which points to hashmap as the winner. We didn't have any perf tests for
fast-export or fast-import, so I added one as a more real-world case.
It uses an export without blobs since that's significantly cheaper than
a full one, but still is an interesting case people might use (e.g., for
rewriting history). It will emphasize this change in some ways (as a
percentage we spend more time making objects and less shuffling blob
bytes around) and less in others (the total object count is lower).
Here are the results for linux.git:
Test HEAD^ HEAD
----------------------------------------------------------------------------
9300.1: export (no-blobs) 67.64(66.96+0.67) 67.81(67.06+0.75) +0.3%
9300.2: import (no-blobs) 284.04(283.34+0.69) 198.09(196.01+0.92) -30.3%
It only has ~5.2M commits and trees, so this is a larger effect than I
expected (the 2^23 case above only improved by 50s or so, but here we
gained almost 90s). This is probably due to actually performing more
object lookups in a real import with trees and commits, as opposed to
just dumping a bunch of blobs into a pack.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 645c432d61 (pack-objects: use reachability bitmap index when
generating non-stdout pack, 2016-09-10) added two timing tests for
packing to an on-disk file, both with and without bitmaps. However, the
non-bitmap one isn't interesting to have as part of p5310's regression
suite. It _could_ be used as a baseline to show off the improvement in
the bitmap case, but:
- the point of the t/perf suite is to find performance regressions,
and it won't help with that. We don't compare the numbers between
two tests (which the perf suite has no idea are even related), and
any change in its numbers would have nothing to do with bitmaps.
- it did show off the improvement in the commit message of 645c432d61,
but it wasn't even necessary there. The bitmap case already shows an
improvement (because before the patch, it behaved the same as the
non-bitmap case), and the perf suite is even able to show the
difference between the before and after measurements.
On top of that, it's one of the most expensive tests in the suite,
clocking in around 60s for linux.git on my machine (as compared to 16s
for the bitmapped version). And by default when using "./run", we'd run
it three times!
So let's just drop it. It's not useful and is adding minutes to perf
runs.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Just as rev-list recently learned to combine filters and bitmaps, let's
do the same for pack-objects. The infrastructure is all there; we just
need to pass along our filter options, and the pack-bitmap code will
decide to use bitmaps or not.
This unsurprisingly makes things faster for partial clones of large
repositories (here we're cloning linux.git):
Test HEAD^ HEAD
------------------------------------------------------------------------------
5310.11: simulated partial clone 38.94(37.28+5.87) 11.06(11.27+4.07) -71.6%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Just as the previous commit implemented BLOB_NONE, we can support
BLOB_LIMIT filters by looking at the sizes of any blobs in the result
and unsetting their bits as appropriate. This is slightly more expensive
than BLOB_NONE, but still produces a noticeable speedup (these results
are on git.git):
Test HEAD~2 HEAD
------------------------------------------------------------------------------------
5310.9: rev-list count with blob:none 1.80(1.77+0.02) 0.22(0.20+0.02) -87.8%
5310.10: rev-list count with blob:limit=1k 1.99(1.96+0.03) 0.29(0.25+0.03) -85.4%
The implementation is similar to the BLOB_NONE one, with the exception
that we have to go object-by-object while walking the blob-type bitmap
(since we can't mask out the matches, but must look up the size
individually for each blob). The trick with using ctz64() is taken from
show_objects_for_type(), which likewise needs to find individual bits
(but wants to quickly skip over big chunks without blobs).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We can easily support BLOB_NONE filters with bitmaps. Since we know the
types of all of the objects, we just need to clear the result bits of
any blobs.
Note two subtleties in the implementation (which I also called out in
comments):
- we have to include any blobs that were specifically asked for (and
not reached through graph traversal) to match the non-bitmap version
- we have to handle in-pack and "ext_index" objects separately.
Arguably prepare_bitmap_walk() could be adding these ext_index
objects to the type bitmaps. But it doesn't for now, so let's match
the rest of the bitmap code here (it probably wouldn't be an
efficiency improvement to do so since the cost of extending those
bitmaps is about the same as our loop here, but it might make the
code a bit simpler).
Here are perf results for the new test on git.git:
Test HEAD^ HEAD
--------------------------------------------------------------------------------
5310.9: rev-list count with blob:none 1.67(1.62+0.05) 0.22(0.21+0.02) -86.8%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ever since we added reachability bitmap support, we've been able to use
it with rev-list to get the full list of objects, like:
git rev-list --objects --use-bitmap-index --all
But you can't do so without --objects, since we weren't ready to just
show the commits. However, the internals of the bitmap code are mostly
ready for this: they avoid opening up trees when walking to fill in the
bitmaps. We just need to actually pass in the rev_info to
traverse_bitmap_commit_list() so it knows which types to bother
triggering our callback for.
For completeness, the perf test now covers both the existing --objects
case, as well as the new commits-only behavior (the objects one got way
faster when we introduced bitmaps, but obviously isn't improved now).
Here are numbers for linux.git:
Test HEAD^ HEAD
------------------------------------------------------------------------
5310.7: rev-list (commits) 8.29(8.10+0.19) 1.76(1.72+0.04) -78.8%
5310.8: rev-list (objects) 8.06(7.94+0.12) 8.14(7.94+0.13) +1.0%
That run was cheating a little, as I didn't have any commit-graph in the
repository, and we'd built it by default these days when running git-gc.
Here are numbers with a commit-graph:
Test HEAD^ HEAD
------------------------------------------------------------------------
5310.7: rev-list (commits) 0.70(0.58+0.12) 0.51(0.46+0.04) -27.1%
5310.8: rev-list (objects) 6.20(6.09+0.10) 6.27(6.16+0.11) +1.1%
Still an improvement, but a lot less impressive.
We could have the perf script remove any commit-graph to show the
out-sized effect, but it probably makes sense to leave it in what would
be a more typical setup.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a repository with many packfiles, the cost of the procedure that
avoids registering the same packfile twice was unnecessarily high
by using an inefficient search algorithm, which has been corrected.
* cs/store-packfiles-in-hashmap:
packfile.c: speed up loading lots of packfiles
PerfTest fix to avoid stale result mixed up with the latest round
of test results.
* tg/perf-remove-stale-result:
perf-lib: use a single filename for all measurement types
When loading packfiles on start-up, we traverse the internal packfile
list once per file to avoid reloading packfiles that have already
been loaded. This check runs in quadratic time, so for poorly
maintained repos with a large number of packfiles, it can be pretty
slow.
Add a hashmap containing the packfile names as we load them so that
the average runtime cost of checking for already-loaded packs becomes
constant.
Add a perf test to p5303 to show speed-up.
The existing p5303 test runtimes are dominated by other factors and do
not show an appreciable speed-up. The new test in p5303 clearly exposes
a speed-up in bad cases. In this test we create 10,000 packfiles and
measure the start-up time of git rev-parse, which does little else
besides load in the packs.
Here are the numbers for the new p5303 test:
Test HEAD^ HEAD
---------------------------------------------------------------------
5303.12: load 10,000 packs 1.03(0.92+0.10) 0.12(0.02+0.09) -88.3%
Signed-off-by: Colin Stolley <cstolley@runbox.com>
Helped-by: Jeff King <peff@peff.net>
[jc: squashed the change to call hashmap in install_packed_git() by peff]
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The perf suite's aggregate.perl depends on Git.pm, which is a mild
annoyance if you've built git with NO_PERL. It turns out that the only
thing we use it for is a single call of the command_oneline() helper.
We can just replace this with backticks or similar.
Annoyingly, perl has no backtick equivalent that avoids a shell eval,
which means our $arg would require quoting. This probably doesn't matter
for our purposes, but it's better to be safe and model good style. So
we'll just provide a short helper around open(), which takes its
arguments as a list.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The perf tests write files recording the results of tests. These
results are later aggregated by 'aggregate.perl'. If the tests are run
multiple times, those results are overwritten by the new results. This
works just fine as long as there are only perf tests measuring the
times, whose results are stored in "$base".times files.
However 22bec79d1a ("t/perf: add infrastructure for measuring sizes",
2018-08-17) introduced a new type of test for measuring the size of
something. The results of this are written to "$base".size files.
"$base" is essentially made up of the basename of the script plus the
test number. So if test numbers shift because a new test was
introduced earlier in the script we might end up with both a ".times"
and a ".size" file for the same test. In the aggregation script the
".times" file is preferred over the ".size" file, so some size tests
might end with performance numbers from a previous run of the test.
This is mainly relevant when writing perf tests that check both
performance and sizes, and can get quite confusing during
developement.
We could fix this by doing a more thorough job of cleaning out old
".times" and ".size" files before running each test. However, an even
easier solution is to just use the same filename for both types of
measurement, meaning we'll always overwrite the previous result. We
don't even need to change the file format to distinguish the two;
aggregate.perl already decides which is which based on a regex of the
content (this may become ambiguous if we add new types in the future,
but we could easily add a header field to the file at that point).
Based on an initial patch from Thomas Gummerer, who discovered the
problem and did all of the analysis (which I stole for the commit
message above):
https://public-inbox.org/git/20191119185047.8550-1-t.gummerer@gmail.com/
Helped-by: Thomas Gummerer <t.gummerer@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This patch fixes an extreme slowdown in pack-objects when you have more
than 1023 packs. See below for numbers.
Since 43fa44fa3b (pack-objects: move in_pack out of struct object_entry,
2018-04-14), we use a complicated system to save some per-object memory.
Each object_entry structs gets a 10-bit field to store the index of the
pack it's in. We map those indices into pointers using
packing_data->in_pack_by_idx, which we initialize at the start of the
program. If we have 2^10 or more packs, then we instead create an array
of pack pointers, one per object. This is packing_data->in_pack.
So far so good. But there's one other tricky case: if a new pack arrives
after we've initialized in_pack_by_idx, it won't have an index yet. We
solve that by calling oe_map_new_pack(), which just switches on the fly
to the less-optimal in_pack mechanism, allocating the array and
back-filling it for already-seen objects.
But that logic kicks in even when we've switched to it already (whether
because we really did see a new pack, or because we had too many packs
in the first place). The result doesn't produce a wrong outcome, but
it's very slow. What happens is this:
- imagine you have a repo with 500k objects and 2000 packs that you
want to repack.
- before looking at any objects, we call prepare_in_pack_by_idx(). It
starts allocating an index for each pack. On the 1024th pack, it
sees there are too many, so it bails, leaving in_pack_by_idx as
NULL.
- while actually adding objects to the packing list, we call
oe_set_in_pack(), which checks whether the pack already has an
index. If it's one of the packs after the first 1023, then it
doesn't have one, and we'll call oe_map_new_pack().
But there's no useful work for that function to do. We're already
using in_pack, so it just uselessly walks over the complete list of
objects, trying to backfill in_pack.
And we end up doing this for almost 1000 packs (each of which may be
triggered by more than one object). And each time it triggers, we
may iterate over up to 500k objects. So in the absolute worst case,
this is quadratic in the number of objects.
The solution is simple: we don't need to bother checking whether the
pack has an index if we've already converted to using in_pack, since by
definition we're not going to use it. So we can just push the "does the
pack have a valid index" check down into that half of the conditional,
where we know we're going to use it.
The current test in p5303 sadly doesn't notice this problem, since it
maxes out at 1000 packs. If we add a new test to it at 2000 packs, it
does show the improvement:
Test HEAD^ HEAD
----------------------------------------------------------------------
5303.12: repack (2000) 26.72(39.68+0.67) 15.70(28.70+0.66) -41.2%
However, these many-pack test cases are rather expensive to run, so
adding larger and larger numbers isn't appealing. Instead, we can show
it off more easily by using GIT_TEST_FULL_IN_PACK_ARRAY, which forces us
into the absolute worst case: no pack has an index, so we'll trigger
oe_map_new_pack() pointlessly for every single object, making it truly
quadratic.
Here are the numbers (on git.git) with the included change to p5303:
Test HEAD^ HEAD
----------------------------------------------------------------------
5303.3: rev-list (1) 2.05(1.98+0.06) 2.06(1.99+0.06) +0.5%
5303.4: repack (1) 33.45(33.46+0.19) 2.75(2.73+0.22) -91.8%
5303.6: rev-list (50) 2.07(2.01+0.06) 2.06(2.01+0.05) -0.5%
5303.7: repack (50) 34.21(35.18+0.16) 3.49(4.50+0.12) -89.8%
5303.9: rev-list (1000) 2.87(2.78+0.08) 2.88(2.80+0.07) +0.3%
5303.10: repack (1000) 41.26(51.30+0.47) 10.75(20.75+0.44) -73.9%
Again, those improvements aren't realistic for the 1-pack case (because
in the real world, the full-array solution doesn't kick in), but it's
more useful to be testing the more-complicated code path.
While we're looking at this issue, we'll tweak one more thing: in
oe_map_new_pack(), we call REALLOC_ARRAY(pack->in_pack). But we'd never
expect to get here unless we're back-filling it for the first time, in
which case it would be NULL. So let's switch that to ALLOC_ARRAY() for
clarity, and add a BUG() to document the expectation. Unfortunately this
code isn't well-covered in the test suite because it's inherently racy
(it only kicks in if somebody else adds a new pack while we're in the
middle of repacking).
Signed-off-by: Jeff King <peff@peff.net>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are two perf scripts numbered p5600, but with otherwise different
names ("clone-reference" versus "partial-clone"). We store timing
results in files named after the whole script, so internally we don't
get confused between the two. But "aggregate.perl" just prints the test
number for each result, giving multiple entries for "5600.3". It also
makes it impossible to skip one test but not the other with
GIT_SKIP_TESTS.
Let's renumber the one that appeared later (by date -- the source of the
problem is that the two were developed on independent branches). For the
non-perf test suite, our test-lint rule would have complained about this
when the two were merged, but t/perf never learned that trick.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we receive a remote ref update to sha1 "X", we want to check that
we have all of the objects needed by "X". We can assume that our
repository is not currently corrupted, and therefore if we have a ref
pointing at "Y", we have all of its objects. So we can stop our
traversal from "X" as soon as we hit "Y".
If we make the same non-corruption assumption about any repositories we
use to store alternates, then we can also use their ref tips to shorten
the traversal.
This is especially useful when cloning with "--reference", as we
otherwise do not have any local refs to check against, and have to
traverse the whole history, even though the other side may have sent us
few or no objects. Here are results for the included perf test (which
shows off more or less the maximal savings, getting one new commit and
sharing the whole history):
Test HEAD^ HEAD
--------------------------------------------------------------------
[on git.git]
5600.3: clone --reference 2.94(2.86+0.08) 0.09(0.08+0.01) -96.9%
[on linux.git]
5600.3: clone --reference 45.74(45.34+0.41) 0.36(0.30+0.08) -99.2%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Performance test framework has been broken and measured the version
of Git that happens to be on $PATH, not the specified one to
measure, for a while, which has been corrected.
* ab/perf-installed-fix:
perf-lib.sh: forbid the use of GIT_TEST_INSTALLED
perf tests: add "bindir" prefix to git tree test results
perf-lib.sh: remove GIT_TEST_INSTALLED from perf-lib.sh
perf-lib.sh: make "./run <revisions>" use the correct gits
perf aggregate: remove GIT_TEST_INSTALLED from --codespeed
perf README: correct docs for 3c8f12c96c regression
The script to aggregate perf result unconditionally depended on
libjson-perl even though it did not have to, which has been
corrected.
* jk/perf-aggregate-wo-libjson:
t/perf: depend on perl JSON only when using --codespeed
Fix index-pack perf test so that the repeated invocations always
run in an empty repository, which emulates the initial clone
situation better.
* jk/p5302-avoid-collision-check-cost:
p5302: create the repo in each index-pack test
The connectivity bitmaps are created by default in bare
repositories now; also the pathname hash-cache is created by
default to avoid making crappy deltas when repacking.
* ew/repack-with-bitmaps-by-default:
pack-objects: default to writing bitmap hash-cache
t5310: correctly remove bitmaps for jgit test
repack: enable bitmaps by default on bare repos
During an initial "git clone --depth=..." partial clone, it is
pointless to spend cycles for a large portion of the connectivity
check that enumerates and skips promisor objects (which by
definition is all objects fetched from the other side). This has
been optimized out.
* js/partial-clone-connectivity-check:
t/perf: add perf script for partial clones
clone: do faster object check for partial clones
As noted in preceding commits setting GIT_TEST_INSTALLED has never
been supported or documented, and as noted in an earlier t/perf/README
change to the extent that it's been documented nobody's notices that
the example hasn't worked since 3c8f12c96c ("test-lib: reorder and
include GIT-BUILD-OPTIONS a lot earlier", 2012-06-24).
We could directly support GIT_TEST_INSTALLED for invocations without
the "run" script, such as:
GIT_TEST_INSTALLED=../../ ./p0000-perf-lib-sanity.sh
GIT_TEST_INSTALLED=/home/avar/g/git ./p0000-perf-lib-sanity.sh
But while not having this "error" will "work", it won't write the the
resulting "test-results/*" files to the right place, and thus a
subsequent call to aggregate.perl won't work as expected.
Let's just tell the user that they need to use the "run" script,
which'll correctly deal with this and set the right
PERF_RESULTS_PREFIX.
If someone's in desperate need of bypassing "run" for whatever reason
they can trivially do so by setting "PERF_SET_GIT_TEST_INSTALLED", but
not we won't have people who expect GIT_TEST_INSTALLED to just work
wondering why their aggregation doesn't work, even though they're
running the right "git".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Change the output file names in test-results/ to be
"test-results/bindir_<munged dir>" rather than just
"test-results/<munged dir>".
This is for consistency with the "build_" directories we have for
built revisions, i.e. "test-results/build_<SHA-1>".
There's no user-visible functional changes here, it just makes it
easier to see at a glance what "test-results" files are of what "type"
as they're all explicitly grouped together now, and to grep this code
to find both the run_dirs_helper() implementation and its
corresponding aggregate.perl code.
Note that we already guarantee that the rest of the
PERF_RESULTS_PREFIX is an absolute path, and since it'll start with
e.g. "/" which we munge to "_" we'll up with a readable string like
"bindir_home_avar[...]".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Follow-up my preceding change which fixed the immediate "./run
<revisions>" regression in 0baf78e7bc ("perf-lib.sh: rely on
test-lib.sh for --tee handling", 2019-03-15) and entirely get rid of
GIT_TEST_INSTALLED from perf-lib.sh (and aggregate.perl).
As noted in that change the dance we're doing with GIT_TEST_INSTALLED
perf-lib.sh isn't necessary, but there I was doing the most minimal
set of changes to quickly fix a regression.
But it's much simpler to never deal with the "GIT_TEST_INSTALLED" we
were setting in perf-lib.sh at all. Instead the run_dirs_helper() sets
the previously inferred $PERF_RESULTS_PREFIX directly.
Setting this at the callsite that's already best positioned to
exhaustively know about all the different cases we need to handle
where PERF_RESULTS_PREFIX isn't what we want already (the empty
string) makes the most sense. In one-off cases like:
./run ./p0000-perf-lib-sanity.sh
./p0000-perf-lib-sanity.sh
We'll just do the right thing because PERF_RESULTS_PREFIX will be
empty, and test-lib.sh takes care of finding where our git is.
Any refactoring of this code needs to change both the shell code and
the Perl code in aggregate.perl, because when running e.g.:
./run ../../ -- <test>
The "../../" path to a relative bindir needs to be munged to a
filename containing the results, and critically aggregate.perl does
not get passed the path to those aggregations, just "../..".
Let's fix cases where aggregate.perl would print e.g. ".." in its
report output for this, and "git" for "/home/avar/g/git", i.e. it
would always pick the last element. Now'll always print the full path
instead.
This also makes the code sturdier, e.g. you can feed "../.." to
"./run" and then an absolute path to the aggregate.perl script, as
long as the absolute path and "../.." resolved to the same directory
printing the aggregation will work.
Also simplify the "[_*]" on the RHS of "tr -c", we're trimming
everything to "_", so we don't need that.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Fix a really bad regression in 0baf78e7bc ("perf-lib.sh: rely on
test-lib.sh for --tee handling", 2019-03-15). Since that change all
runs of different <revisions> of git have used the git found in the
user's $PATH, e.g. /usr/bin/git instead of the <revision> we just
built and wanted to performance test.
The problem starts with GIT_TEST_INSTALLED not working like our
non-perf tests with the "run" script. I.e. you can't run performance
tests against a given installed git. Instead we expect to use it
ourselves to point GIT_TEST_INSTALLED to the <revision> we just built.
However, we had been relying on '$(cd "$GIT_TEST_INSTALLED" && pwd)'
to resolve that relative $GIT_TEST_INSTALLED to an absolute
path *before* test-lib.sh was loaded, in cases where it was
e.g. "build/<rev>/bin-wrappers" and we wanted "<abs_path>build/...".
This change post-dates another proposed solution by a few days[1], I
didn't notice that version when I initially wrote this. I'm doing the
most minimal thing to solve the regression here, a follow-up change
will move this result prefix selection logic entirely into the "run"
script.
This makes e.g. these cases all work:
./run . $PWD/../../ origin/master origin/next HEAD -- <tests>
As well as just a plain one-off:
./run <tests>
And, since we're passing down the new GIT_PERF_DIR_MYDIR_REL we make
sure the bug relating to aggregate.perl not finding our files as
described in 0baf78e7bc doesn't happen again.
What *doesn't* work is setting GIT_TEST_INSTALLED to a relative path,
this will subtly fail in test-lib.sh. This has always been the case
even before 0baf78e7bc, and as documented in t/README the
GIT_TEST_INSTALLED variable should be set to an absolute path (needs
to be set "to the bindir", which is always absolute), and the "perf"
framework expects to munge it itself.
Perhaps that should be dealt with in the future to allow manually
setting GIT_TEST_INSTALLED, but as a preceding commit showed the user
can just use the "run" script, which'll also pick the right output
directory for the test results as expected by aggregate.perl.
1. https://public-inbox.org/git/20190502222409.GA15631@sigill.intra.peff.net/
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Remove the setting of the "environment" from the --codespeed output. I
don't think this is useful, and it helps with a later refactoring
where we GIT_TEST_INSTALLED stop munging/reading GIT_TEST_INSTALLED in
the perf tests in so many places.
This was added in 05eb1c37ed ("perf/aggregate: implement codespeed
JSON output", 2018-01-05), but since the "run" scripts uses
"GIT_TEST_INSTALLED" internally this was only ever useful for one-off
runs of a single revision as all the "environment" values would be
ones for whatever directory the "run" script ran last.
Let's instead fall back on the "uname -r" case, which is the sort of
thing the environment should be set to, not something that duplicates
other parts of the codpseed output. For setting the "environment" to
something custom the perf.repoName variable can be used. See
19cf57a92e ("perf/run: read GIT_PERF_REPO_NAME from perf.repoName",
2018-01-05).
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Since 3c8f12c96c ("test-lib: reorder and include GIT-BUILD-OPTIONS a
lot earlier", 2012-06-24) the suggested advice of overriding
GIT_BUILD_DIR has not worked. We've printed a hard error like this
given e.g. GIT_BUILD_DIR=/home/avar/g/git:
/bin-wrappers/git is not executable; using GIT_EXEC_PATH
error: You haven't built things yet, have you?
Let's just suggest that the user run other gits via the "run"
script. That'll do the right thing for setting the path to the other
git, and running the "aggregate.perl" scripts afterwards will work.
As an aside, if setting GIT_BUILD_DIR had still worked, then the
MODERN_GIT feature/fix added in 1a0962dee5 ("t/perf: fix regression in
testing older versions of git", 2016-06-22) would have broke.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
We don't cover the partial clone feature at all in t/perf. Let's at
least run a few basic tests so that we'll notice any regressions.
We'll do a no-blob clone, and split it into two parts: the actual object
transfer, and the subsequent checkout (which will of course require
another transfer to get the blobs). That will help us more clearly
assess the performance of each.
There are obviously a lot more possibilities besides just a no-blob
partial clone, but this should serve as a canary that alerts us to any
generic slow-downs (and we can add more tests later for cases that
aren't exercised here).
There are a few non-ideal things here that make this not an entirely
accurate test, but are probably OK for our purposes:
1. We have to do some extra prep/cleanup work inside the timing tests,
since they impact the on-disk state and the perf harness may run
each one multiple times.
In practice this is probably OK, since these bits should be much
less expensive than the operations we are measuring.
2. The clone time is likely to be dominated by the server's object
enumeration. In the real world, a repo large enough to drive people
to partial clones is likely to have reachability bitmaps enabled.
And in the opposite direction, our object transfer is happening at
the speed of a local pipe, whereas in the real world it would
bottle-neck on the network.
So any percentage speedups should be taken with a grain of salt.
But hopefully any regressions will produce enough of an effect to
be noticeable.
This script also demonstrates the recent improvement from dfa33a298d
(clone: do faster object check for partial clones, 2019-04-19):
Test dfa33a298d^ dfa33a298d
-------------------------------------------------------------------------
5600.2: clone without blobs 18.41(22.72+1.09) 6.83(11.65+0.50) -62.9%
5600.3: checkout of result 1.82(3.24+0.26) 1.84(3.24+0.26) +1.1%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Performance fix for "rev-list --parents -- pathspec".
* jk/revision-rewritten-parents-in-prio-queue:
revision: use a prio_queue to hold rewritten parents
Commit 05eb1c37ed (perf/aggregate: implement codespeed JSON output,
2018-01-05) added a dependency on the perl JSON module to show output
from aggregate.perl, but we only need it when the user asks for
--codespeed output. While the module is pretty common, it's not part of
the base system, and this dependency can get in the way of producing the
default human-readable output.
Let's bump the "use" down to a "require" in the code path that needs it,
which will be interpreted at run-time instead of compile-time. People
not using "--codespeed" won't even load the module, and anybody using it
should see the same results (including the same perl error if they don't
have it).
Note that this skips the importing step, so we'll have to fully qualify
our function call. We could accomplish the same thing in other ways.
E.g., calling JSON->import() ourselves, or wrapping "use JSON" in an
eval. Since there's only one such call, this seems like the
least-magical way of doing it.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The p5302 script runs "index-pack --stdin" in each timing test. It does
two things to try to get good timings:
1. we do the repo creation in a separate (non-timed) setup test, so
that our timing is purely the index-pack run
2. we use a separate repo for each test; this is important because the
presence of existing objects in the repo influences the result
(because we'll end up doing collision checks against them)
But this forgets one thing: we generally run each timed test multiple
times to reduce the impact of noise. Which means that repeats of each
test after the first will be subject to the collision slowdown from
point 2, and we'll generally just end up taking the first time anyway.
Instead, let's create the repo in the test (effectively undoing point
1). That does add a constant amount of extra work to each iteration, but
it's quite small compared to the actual effects we're interested in
measuring.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This patch fixes a quadratic list insertion in rewrite_one() when
pathspec limiting is combined with --parents. What happens is something
like this:
1. We see that some commit X touches the path, so we try to rewrite
its parents.
2. rewrite_one() loops forever, rewriting parents, until it finds a
relevant parent (or hits the root and decides there are none). The
heavy lifting is done by process_parent(), which uses
try_to_simplify_commit() to drop parents.
3. process_parent() puts any intermediate parents into the
&revs->commits list, inserting by commit date as usual.
So if commit X is recent, and then there's a large chunk of history that
doesn't touch the path, we may add a lot of commits to &revs->commits.
And insertion by commit date is O(n) in the worst case, making the whole
thing quadratic.
We tried to deal with this long ago in fce87ae538 (Fix quadratic
performance in rewrite_one., 2008-07-12). In that scheme, we cache the
oldest commit in the list; if the new commit to be added is older, we
can start our linear traversal there. This often works well in practice
because parents are older than their descendants, and thus we tend to
add older and older commits as we traverse.
But this isn't guaranteed, and in fact there's a simple case where it is
not: merges. Imagine we look at the first parent of a merge and see a
very old commit (let's say 3 years old). And on the second parent, as we
go back 3 years in history, we might have many commits. That one
first-parent commit has polluted our oldest-commit cache; it will remain
the oldest while we traverse a huge chunk of history, during which we
have to fall back to the slow, linear method of adding to the list.
Naively, one might imagine that instead of caching the oldest commit,
we'd start at the last-added one. But that just makes some cases faster
while making others slower (and indeed, while it made a real-world test
case much faster, it does quite poorly in the perf test include here).
Fundamentally, these are just heuristics; our worst case is still
quadratic, and some cases will approach that.
Instead, let's use a data structure with better worst-case performance.
Swapping out revs->commits for something else would have repercussions
all over the code base, but we can take advantage of one fact: for the
rewrite_one() case, nobody actually needs to see those commits in
revs->commits until we've finished generating the whole list.
That leaves us with two obvious options:
1. We can generate the list _unordered_, which should be O(n), and
then sort it afterwards, which would be O(n log n) total. This is
"sort-after" below.
2. We can insert the commits into a separate data structure, like a
priority queue. This is "prio-queue" below.
I expected that sort-after would be the fastest (since it saves us the
extra step of copying the items into the linked list), but surprisingly
the prio-queue seems to be a bit faster.
Here are timings for the new p0001.6 for all three techniques across a
few repositories, as compared to master:
master cache-last sort-after prio-queue
--------------------------------------------------------------------------------------------
GIT_PERF_REPO=git.git
0.52(0.50+0.02) 0.53(0.51+0.02) +1.9% 0.37(0.33+0.03) -28.8% 0.37(0.32+0.04) -28.8%
GIT_PERF_REPO=linux.git
20.81(20.74+0.07) 20.31(20.24+0.07) -2.4% 0.94(0.86+0.07) -95.5% 0.91(0.82+0.09) -95.6%
GIT_PERF_REPO=llvm-project.git
83.67(83.57+0.09) 4.23(4.15+0.08) -94.9% 3.21(3.15+0.06) -96.2% 2.98(2.91+0.07) -96.4%
A few items to note:
- the cache-list tweak does improve the bad case for llvm-project.git
that started my digging into this problem. But it performs terribly
on linux.git, barely helping at all.
- the sort-after and prio-queue techniques work well. They approach
the timing for running without --parents at all, which is what you'd
expect (see below for more data).
- prio-queue just barely outperforms sort-after. As I said, I'm not
really sure why this is the case, but it is. You can see it even
more prominently in this real-world case on llvm-project.git:
git rev-list --parents 07ef786652e7 -- llvm/test/CodeGen/Generic/bswap.ll
where prio-queue routinely outperforms sort-after by about 7%. One
guess is that the prio-queue may just be more efficient because it
uses a compact array.
There are three new perf tests:
- "rev-list --parents" gives us a baseline for running with --parents.
This isn't sped up meaningfully here, because the bad case is
triggered only with simplification. But it's good to make sure we
don't screw it up (now, or in the future).
- "rev-list -- dummy" gives us a baseline for just traversing with
pathspec limiting. This gives a lower bound for the next test (and
it's also a good thing for us to be checking in general for
regressions, since we don't seem to have any existing tests).
- "rev-list --parents -- dummy" shows off the problem (and our fix)
Here are the timings for those three on llvm-project.git, before and
after the fix:
Test master prio-queue
------------------------------------------------------------------------------
0001.3: rev-list --parents 2.24(2.12+0.12) 2.22(2.11+0.11) -0.9%
0001.5: rev-list -- dummy 2.89(2.82+0.07) 2.92(2.89+0.03) +1.0%
0001.6: rev-list --parents -- dummy 83.67(83.57+0.09) 2.98(2.91+0.07) -96.4%
Changes in the first two are basically noise, and you can see we
approach our lower bound in the final one.
Note that we can't fully get rid of the list argument from
process_parents(). Other callers do have lists, and it would be hard to
convert them. They also don't seem to have this problem (probably
because they actually remove items from the list as they loop, meaning
it doesn't grow so large in the first place). So this basically just
drops the "cache_ptr" parameter (which was used only by the one caller
we're fixing here) and replaces it with a prio_queue. Callers are free
to use either data structure, depending on what they're prepared to
handle.
Reported-by: Björn Pettersson A <bjorn.a.pettersson@ericsson.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since its inception, the perf-lib.sh script has manually handled the
"--tee" option (and other options which imply it, like "--valgrind")
with a cut-and-pasted block from test-lib.sh. That block has grown stale
over the years, and has at least three problems:
1. It uses $SHELL to re-exec the script, whereas the version in
test-lib.sh learned to use $TEST_SHELL_PATH.
2. It does an ad-hoc search of the "$*" string, whereas test-lib.sh
learned to carefully parse the arguments left to right.
3. It never learned about --verbose-log (which also implies --tee),
so it would not trigger for that option.
This last one was especially annoying, because t/perf/run uses the
GIT_TEST_OPTS from your config.mak to run the perf scripts. So if you've
set, say, "-x --verbose-log" there, it will be passed as part of most
perf runs. And while this script doesn't recognize the option, the
test-lib.sh that we source _does_, and the behavior ends up being much
more annoying:
- as the comment at the top of the block says, we have to run this
tee code early, before we start munging variables (it says
GIT_BUILD_DIR, but the problematic variable is actually
GIT_TEST_INSTALLED).
- since we don't recognize --verbose-log, we don't trigger the block.
We go on to munge GIT_TEST_INSTALLED, converting it from a relative
to an absolute path.
- then we source test-lib.sh, which _does_ recognize --verbose-log. It
re-execs the script, which runs again. But this time with an
absolute version of GIT_TEST_INSTALLED.
- As a result, we copy the absolute version of GIT_TEST_INSTALLED into
perf_results_prefix. Instead of writing our results to the expected
"test-results/build_1234abcd.p1234-whatever.times", we instead write
them to "test-results/_full_path_to_repo_t_perf_build_1234...".
The aggregate.perl script doesn't expect this, and so it prints
"<missing>" for each result (even though it spent considerable time
running the tests!).
We can solve all of these in one blow by just deleting our custom
handling, and relying on the inclusion of test-lib.sh to handle --tee,
--verbose-log, etc.
There's one catch, though. We want to handle GIT_TEST_INSTALLED after
we've included test-lib.sh, since we want it un-munged in the re-exec'd
version of the script. But if we want to convert it from a relative
to an absolute path, we must do so before we load test-lib.sh, since it
will change our working directory. So we compute the absolute directory
first, store it away, then include test-lib.sh, and finally assign to
GIT_TEST_INSTALLED as appropriate.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Enabling pack.writebitmaphashcache should always be a performance win.
It costs only 4 bytes per object on disk, and the timings in ae4f07fbcc
(pack-bitmap: implement optional name_hash cache, 2013-12-21) show it
improving fetch and partial-bitmap clone times by 40-50%.
The only reason we didn't enable it by default at the time is that early
versions of JGit's bitmap reader complained about the presence of
optional header bits it didn't understand. But that was changed in
JGit's d2fa3987a (Use bitcheck to check for presence of OPT_FULL option,
2013-10-30), which made it into JGit v3.5.0 in late 2014.
So let's turn this option on by default. It's backwards-compatible with
all versions of Git, and if you are also using JGit on the same
repository, you'd only run into problems using a version that's almost 5
years old.
We'll drop the manual setting from all of our test scripts, including
perf tests. This isn't strictly necessary, but it has two advantages:
1. If the hash-cache ever stops being enabled by default, our perf
regression tests will notice.
2. We can use the modified perf tests to show off the behavior of an
otherwise unconfigured repo, as shown below.
These are the results of a few of a perf tests against linux.git that
showed interesting results. You can see the expected speedup in 5310.4,
which was noted in ae4f07fbcc. Curiously, 5310.8 did not improve (and
actually got slower), despite seeing the opposite in ae4f07fbcc.
I don't have an explanation for that.
The tests from p5311 did not exist back then, but do show improvements
(a smaller pack due to better deltas, which we found in less time).
Test HEAD^ HEAD
-------------------------------------------------------------------------------------
5310.4: simulated fetch 7.39(22.70+0.25) 5.64(11.43+0.22) -23.7%
5310.8: clone (partial bitmap) 18.45(24.83+1.19) 19.94(28.40+1.36) +8.1%
5311.31: server (128 days) 0.41(1.13+0.05) 0.34(0.72+0.02) -17.1%
5311.32: size (128 days) 7.4M 7.0M -4.8%
5311.33: client (128 days) 1.33(1.49+0.06) 1.29(1.37+0.12) -3.0%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
