Re: [PATCH] hash: reduce size of algo member of object ID

Am 04.10.21 um 10:20 schrieb Jeff King:

On Mon, Oct 04, 2021 at 04:13:34AM -0400, Jeff King wrote:

quoted

It looks like adding the "algo" field did make a big difference for the
oid_array case, but changing it to a char doesn't seem to help at all:

  $ hyperfine -L v none,int,char './git.{v} cat-file --batch-all-objects --batch-check="%(objectname)"'
  Benchmark #1: ./git.none cat-file --batch-all-objects --batch-check="%(objectname)"
    Time (mean ± σ):      1.653 s ±  0.009 s    [User: 1.607 s, System: 0.046 s]
    Range (min … max):    1.640 s …  1.670 s    10 runs

  Benchmark #2: ./git.int cat-file --batch-all-objects --batch-check="%(objectname)"
    Time (mean ± σ):      1.067 s ±  0.012 s    [User: 1.017 s, System: 0.050 s]
    Range (min … max):    1.053 s …  1.089 s    10 runs

  Benchmark #3: ./git.char cat-file --batch-all-objects --batch-check="%(objectname)"
    Time (mean ± σ):      1.092 s ±  0.013 s    [User: 1.046 s, System: 0.046 s]
    Range (min … max):    1.080 s …  1.116 s    10 runs

  Summary
    './git.int cat-file --batch-all-objects --batch-check="%(objectname)"' ran
      1.02 ± 0.02 times faster than './git.char cat-file --batch-all-objects --batch-check="%(objectname)"'
      1.55 ± 0.02 times faster than './git.none cat-file --batch-all-objects --batch-check="%(objectname)"'

I'm actually surprised it had this much of an impact. But I guess this
benchmark really is mostly just memcpy-ing oids into a big array,
sorting it, and printing the result. If that array is 12% bigger, we'd
expect at least a 12% speedup. But adding in non-linear elements like
growing the array (though I guess that is amortized linear) and sorting
(which picks up an extra log(n) term) make the difference.

It's _kind of_ silly in a sense, since usually you'd ask for other parts
of the object, which will make the speed difference relatively smaller.
But just dumping a bunch of oids is actually not an unreasonable thing
to do. I suspect it got a lot slower with 32-byte GIT_MAX_RAWSZ, too
(even when you're using 20-byte sha1), but I don't think there's an easy
way to get out of that.

Oh wait, I'm reading it totally wrong. Adding in the extra 4 bytes
actually made it _faster_ than not having an algo field. Now I'm
super-confused. I could believe that it gave us some better alignment,
but the original struct was 32 bytes. 36 seems like a strictly worse
number.

Apple clang 13 on arm64-apple-darwin20.6.0:

cf0983213c^ (no algo):
Benchmark #1: ./git -C ../linux cat-file --batch-all-objects --batch-check="%(objectname)"
  Time (mean ± σ):      1.105 s ±  0.017 s    [User: 1.058 s, System: 0.045 s]
  Range (min … max):    1.086 s …  1.126 s    10 runs

cf0983213c (int algo):
Benchmark #1: ./git -C ../linux cat-file --batch-all-objects --batch-check="%(objectname)"
  Time (mean ± σ):      1.186 s ±  0.015 s    [User: 1.134 s, System: 0.051 s]
  Range (min … max):    1.162 s …  1.201 s    10 runs

cf0983213c (unsigned char algo):
Benchmark #1: ./git -C ../linux cat-file --batch-all-objects --batch-check="%(objectname)"
  Time (mean ± σ):      1.409 s ±  0.013 s    [User: 1.367 s, System: 0.041 s]
  Range (min … max):    1.397 s …  1.429 s    10 runs


current master (int algo):
Benchmark #1: ./git -C ../linux cat-file --batch-all-objects --batch-check="%(objectname)"
  Time (mean ± σ):      1.217 s ±  0.018 s    [User: 1.170 s, System: 0.046 s]
  Range (min … max):    1.202 s …  1.246 s    10 runs

current master with the patch (unsigned char algo):
Benchmark #1: ./git -C ../linux cat-file --batch-all-objects --batch-check="%(objectname)"
  Time (mean ± σ):      1.465 s ±  0.012 s    [User: 1.417 s, System: 0.046 s]
  Range (min … max):    1.449 s …  1.480 s    10 runs

OK, I don't like my patch anymore.  Weird, though.

René