Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
merge-ort: begin performance work; instrument with trace2_region_* calls
Add some timing instrumentation for both merge-ort and diffcore-rename;
I used these to measure and optimize performance in both, and several
future patch series will build on these to reduce the timings of some
select testcases.
=== Setup ===
The primary testcase I used involved rebasing a random topic in the
linux kernel (consisting of 35 patches) against an older version. I
added two variants, one where I rename a toplevel directory, and another
where I only rebase one patch instead of the whole topic. The setup is
as follows:
$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git
$ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e
$ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34
$ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e
$ git switch -c 5.4-renames v5.4
$ git mv drivers pilots # Introduce over 26,000 renames
$ git commit -m "Rename drivers/ to pilots/"
$ git config merge.renameLimit 30000
$ git config merge.directoryRenames true
=== Testcases ===
Now with REBASE standing for either "git rebase [--merge]" (using
merge-recursive) or "test-tool fast-rebase" (using merge-ort), the
testcases are:
Testcase #1: no-renames
$ git checkout v5.4^0
$ REBASE --onto HEAD base hwmon-updates
Note: technically the name is misleading; there are some renames, but
very few. Rename detection only takes about half the overall time.
Testcase #2: mega-renames
$ git checkout 5.4-renames^0
$ REBASE --onto HEAD base hwmon-updates
Testcase #3: just-one-mega
$ git checkout 5.4-renames^0
$ REBASE --onto HEAD base hwmon-just-one
=== Timing results ===
Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames,
10 runs for the other two cases):
merge-recursive merge-ort
no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s
mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s
just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s
A single re-run of each with some breakdowns:
--- no-renames ---
merge-recursive merge-ort
overall runtime: 19.302 s 14.257 s
inexact rename detection: 7.603 s 7.906 s
everything else: 11.699 s 6.351 s
--- mega-renames ---
merge-recursive merge-ort
overall runtime: 5950.195 s 5499.672 s
inexact rename detection: 5746.309 s 5487.120 s
everything else: 203.886 s 17.552 s
--- just-one-mega ---
merge-recursive merge-ort
overall runtime: 151.001 s 158.582 s
inexact rename detection: 143.448 s 157.835 s
everything else: 7.553 s 0.747 s
=== Timing observations ===
0) Maximum speedup
The "everything else" row represents the maximum speedup we could
achieve if we were to somehow infinitely parallelize inexact rename
detection, but leave everything else alone. The fact that this is so
much smaller than the real runtime (even in the case with virtually no
renames) makes it clear just how overwhelmingly large the time spent on
rename detection can be.
1) no-renames
1a) merge-ort is faster than merge-recursive, which is nice. However,
this still should not be considered good enough. Although the "merge"
backend to rebase (merge-recursive) is sometimes faster than the "apply"
backend, this is one of those cases where it is not. In fact, even
merge-ort is slower. The "apply" backend can complete this testcase in
6.940 s ± 0.485 s
which is about 2x faster than merge-ort and 3x faster than
merge-recursive. One goal of the merge-ort performance work will be to
make it faster than git-am on this (and similar) testcases.
2) mega-renames
2a) Obviously rename detection is a huge cost; it's where most the time
is spent. We need to cut that down. If we could somehow infinitely
parallelize it and drive its time to 0, the merge-recursive time would
drop to about 204s, and the merge-ort time would drop to about 17s. I
think this particular stat shows I've subtly baked a couple performance
improvements into merge-ort and into fast-rebase already.
3) just-one-mega
3a) not much to say here, it just gives some flavor for how rebasing
only one patch compares to rebasing 35.
=== Goals ===
This patch is obviously just the beginning. Here are some of my goals
that this measurement will help us achieve:
* Drive the cost of rename detection down considerably for merges
* After the above has been achieved, see if there are other slowness
factors (which would have previously been overshadowed by rename
detection costs) which we can then focus on and also optimize.
* Ensure our rebase testcase that requires little rename detection
is noticeably faster with merge-ort than with apply-based rebase.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Taylor Blau <ttaylorr@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>- Loading branch information
