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>
In p0001, a variable was created in a test_expect_success block to be
used in later test_perf blocks, but was not exported. This caused the
variable to not appear in those blocks (this can be verified by writing
'test -n "$commit"' in those blocks), resulting in a slightly different
invocation than what was intended. Export that variable.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We time a straight "rev-list --all" and its "--object"
counterpart, both going all the way to the root. However, we
do not time a partial history walk. This patch adds an
extreme case: a walk over a very small slice of history, but
with a very large set of UNINTERESTING tips. This is similar
to the connectivity check run by git on a small fetch, or
the walk done by any pre-receive hooks that want to check
incoming commits.
This test reveals a performance regression in git v1.8.4.2,
caused by fbd4a70 (list-objects: mark more commits as edges
in mark_edges_uninteresting, 2013-08-16):
Test fbd4a703^ fbd4a703
------------------------------------------------------------------------------------------
0001.1: rev-list --all 0.69(0.67+0.02) 0.69(0.68+0.01) +0.0%
0001.2: rev-list --all --objects 3.47(3.44+0.02) 3.48(3.44+0.03) +0.3%
0001.4: rev-list $commit --not --all 0.04(0.04+0.00) 0.04(0.04+0.00) +0.0%
0001.5: rev-list --objects $commit --not --all 0.04(0.03+0.00) 0.27(0.24+0.02) +575.0%
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This introduces a performance testing framework under t/perf/. It
tries to be as close to the test-lib.sh infrastructure as possible,
and thus should be easy to get used to for git developers.
The following points were considered for the implementation:
1. You usually want to compare arbitrary revisions/build trees against
each other. They may not have the performance test under
consideration, or even the perf-lib.sh infrastructure.
To cope with this, the 'run' script lets you specify arbitrary
build dirs and revisions. It even automatically builds the revisions
if it doesn't have them at hand yet.
2. Usually you would not want to run all tests. It would take too
long anyway. The 'run' script lets you specify which tests to run;
or you can also do it manually. There is a Makefile for
discoverability and 'make clean', but it is not meant for
real-world use.
3. Creating test repos from scratch in every test is extremely
time-consuming, and shipping or downloading such large/weird repos
is out of the question.
We leave this decision to the user. Two different sizes of test
repos can be configured, and the scripts just copy one or more of
those (using hardlinks for the object store). By default it tries
to use the build tree's git.git repository.
This is fairly fast and versatile. Using a copy instead of a clone
preserves many properties that the user may want to test for, such
as lots of loose objects, unpacked refs, etc.
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King
Jonathan Tan
Thomas Rast