Re: [PATCH v2 0/7] Sparse Index: integrate with reset

[Victoria Dye <vdye@github.com>]

Ævar Arnfjörð Bjarmason wrote:
> 
> On Tue, Oct 05 2021, Victoria Dye via GitGitGadget wrote:
> 
>> The p2000 tests demonstrate an overall ~70% execution time reduction across
>> all tested usages of git reset using a sparse index:
> 
> [...]
> 
>> Test                                               before   after       
>> ------------------------------------------------------------------------
>> 2000.22: git reset (full-v3)                       0.48     0.51 +6.3% 
>> 2000.23: git reset (full-v4)                       0.47     0.50 +6.4% 
>> 2000.24: git reset (sparse-v3)                     0.93     0.30 -67.7%
>> 2000.25: git reset (sparse-v4)                     0.94     0.29 -69.1%
>> 2000.26: git reset --hard (full-v3)                0.69     0.68 -1.4% 
>> 2000.27: git reset --hard (full-v4)                0.75     0.68 -9.3% 
>> 2000.28: git reset --hard (sparse-v3)              1.29     0.34 -73.6%
>> 2000.29: git reset --hard (sparse-v4)              1.31     0.34 -74.0%
>> 2000.30: git reset -- does-not-exist (full-v3)     0.54     0.51 -5.6% 
>> 2000.31: git reset -- does-not-exist (full-v4)     0.54     0.52 -3.7% 
>> 2000.32: git reset -- does-not-exist (sparse-v3)   1.02     0.31 -69.6%
>> 2000.33: git reset -- does-not-exist (sparse-v4)   1.07     0.30 -72.0%
> 
> This series looks like it really improves some cases, but at the cost of
> that -70% improvement we've got a ~5% regression in 7/7 for the full-v3
> --does-not-exist cases. As noted in your 7/7 (which improves all other
> cases):
> 
>     (full-v3)     0.79(0.38+0.30)   0.91(0.43+0.34) +15.2%
>     (full-v4)     0.80(0.38+0.29)   0.85(0.40+0.35) +6.2%
> 

New performance numbers at the end - I think I have an explanation for this.

> Which b.t.w. I had to read a couple of times before realizig that its
> quoted:
>     
>     Test          before            after
>     ------------------------------------------------------
>     (full-v3)     0.79(0.38+0.30)   0.91(0.43+0.34) +15.2%
>     (full-v4)     0.80(0.38+0.29)   0.85(0.40+0.35) +6.2%
>     (sparse-v3)   0.76(0.43+0.69)   0.44(0.08+0.67) -42.1%
>     (sparse-v4)   0.71(0.40+0.65)   0.41(0.09+0.65) -42.3%
> 
> Is just the does-not-exist part of this bigger table, are the other
> cases all ~0% changed, or ...?
>     

These numbers were for the `git reset -- does-not-exist` case only. If I end
up needing to send a V3, though, I'll probably remove the performance
numbers from 7/7 altogether - looking at them now, they make the commit
message somewhat cluttered. That said, performance numbers *are* helpful for
reviews on the mailing list, so I'd keep the information in the cover
letter at the very least.

> Anyway, until 7/7 the v3 had been sped up, but a ~10% increase landed us
> at ~+6%, and full-v4 had been ~0% but got ~6% worse?
> 
> Is there a way we can get those improvements in performance without
> regressing on the full-* cases?
> 
> Also, these tests only check sparse performance, but isn't some of the
> code being modified here general enough to not be used exclusively by
> the sparse mode, full checkout cone or not?
> 
> It looks fairly easy to extend p2000-sparse-operations.sh to run the
> same tests but just pretend that it's running in a "full" mode without
> actually setting up anyting sparse-specific (the meat of those tests
> just runs "git status" etc. How does that look with this series?
> 

I updated `p2000` locally to do this but the setup was substantially slower
for the full checkout, to the point that it was infeasible to run the
complete test for all relevant commits. Looking at the changes in this
series, nothing appears to affect the full checkout case differently than
the sparse checkout/full index case, so I'm fairly confident there won't be
a regression specific to full checkouts.

> Since only the CL and 7/7 quote numbers from p2000, and 7/7 is at least
> a partial regression, it would be nice to have perf numbers on each
> commit (if only as a one-off for ML consumption). Are there any more
> improvements followed by regressions followed by improvements as we go
> along? Would be useful to know...
> 

I don't think any of the apparent slowdowns seen in these results represent
real regressions. After re-running the performance tests, I saw variability
of up to ~20% execution time across changes with commands that should see no
effect on their execution time (e.g. sparse-v* from 1/7 to 4/7).
Additionally, I saw different increases & decreases each time for each
end-to-end run of the tests. The most reliable, noticeable changes across
the test executions were:

1. When each variant of `git reset` was integrated with sparse index, a
   65-75% execution time reduction in relevant sparse-v* tests.
2. `git reset -- does-not-exist` slower than `git reset` in 6/7,
   then matching its speed after 7/7.
3. As of 7/7, full-v* to sparse-v* showing a 50% execution time reduction.

My guess is that the variability comes from general "uncontrolled" factors
when running the tests (e.g., background processes on my system). The good
news is, when the tests are re-run with more trials (and the recent bugfix
to `t/perf/perf-lib.sh` [1]), the execution times look a lot less worrisome 
(apologies for the table width, but I'd like to err on the side of providing
more complete information):

Test                                               base              [1/7]                    [4/7]                    [5/7]                    [6/7]                    [7/7]            
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
2000.22: git reset (full-v3)                       0.44(0.16+0.19)   0.44(0.17+0.18) +0.0%    0.44(0.17+0.19) +0.0%    0.45(0.17+0.18) +2.3%    0.44(0.17+0.19) +0.0%    0.45(0.17+0.18) +2.3% 
2000.23: git reset (full-v4)                       0.43(0.16+0.18)   0.43(0.16+0.19) +0.0%    0.45(0.17+0.18) +4.7%    0.44(0.17+0.18) +2.3%    0.44(0.17+0.18) +2.3%    0.44(0.18+0.18) +2.3% 
2000.24: git reset (sparse-v3)                     0.82(0.54+0.19)   0.84(0.56+0.19) +2.4%    0.81(0.54+0.19) -1.2%    0.88(0.60+0.19) +7.3%    0.27(0.03+0.45) -67.1%   0.27(0.03+0.47) -67.1%
2000.25: git reset (sparse-v4)                     0.82(0.55+0.18)   0.82(0.53+0.20) +0.0%    0.83(0.55+0.19) +1.2%    0.82(0.54+0.19) +0.0%    0.27(0.03+0.50) -67.1%   0.27(0.03+0.48) -67.1%
2000.26: git reset --hard (full-v3)                0.71(0.38+0.24)   0.69(0.37+0.23) -2.8%    0.70(0.37+0.24) -1.4%    0.78(0.41+0.27) +9.9%    0.71(0.38+0.25) +0.0%    0.70(0.37+0.23) -1.4% 
2000.27: git reset --hard (full-v4)                0.71(0.38+0.23)   0.77(0.42+0.25) +8.5%    0.76(0.41+0.26) +7.0%    0.72(0.40+0.24) +1.4%    0.68(0.37+0.23) -4.2%    0.67(0.36+0.22) -5.6%
2000.28: git reset --hard (sparse-v3)              1.29(0.93+0.26)   1.33(0.95+0.27) +3.1%    1.11(0.76+0.25) -14.0%   0.38(0.05+0.25) -70.5%   0.36(0.04+0.22) -72.1%   0.34(0.04+0.21) -73.6%
2000.29: git reset --hard (sparse-v4)              1.17(0.84+0.24)   1.10(0.79+0.23) -6.0%    1.01(0.69+0.24) -13.7%   0.42(0.05+0.26) -64.1%   0.39(0.05+0.25) -66.7%   0.38(0.05+0.23) -67.5%
2000.30: git reset -- does-not-exist (full-v3)     0.50(0.19+0.20)   0.50(0.19+0.20) +0.0%    0.53(0.21+0.22) +6.0%    0.47(0.18+0.19) -6.0%    0.45(0.18+0.18) -10.0%   0.45(0.18+0.19) -10.0%
2000.31: git reset -- does-not-exist (full-v4)     0.45(0.18+0.18)   0.46(0.18+0.19) +2.2%    0.47(0.19+0.19) +4.4%    0.45(0.18+0.19) +0.0%    0.45(0.18+0.18) +0.0%    0.45(0.18+0.18) +0.0% 
2000.32: git reset -- does-not-exist (sparse-v3)   1.01(0.70+0.21)   0.91(0.62+0.20) -9.9%    0.93(0.64+0.20) -7.9%    0.89(0.61+0.20) -11.9%   0.48(0.23+0.46) -52.5%   0.27(0.03+0.49) -73.3%
2000.33: git reset -- does-not-exist (sparse-v4)   0.99(0.67+0.21)   1.02(0.70+0.22) +3.0%    1.04(0.70+0.22) +5.1%    0.83(0.55+0.19) -16.2%   0.48(0.24+0.48) -51.5%   0.27(0.03+0.49) -72.7%

Note that some commits in this series are not included because they don't
touch any code used by `git reset`.

[1] https://lore.kernel.org/git/pull.1051.git.1633386543759.gitgitgadget@gmail.com/