On Fri, Aug 07, 2020 at 05:32:39AM -0400, Jeff King wrote:
> On Fri, Aug 07, 2020 at 11:04:12AM +0200, Patrick Steinhardt wrote:
> 
> > > It's perhaps more justifiable when there isn't a hook (we're still just
> > > paying that one access(), but it's proportionally bigger). I kind of
> > > doubt it's measurable, though, since a ref write requires a bunch of
> > > syscalls anyway.
> > 
> > Yeah, this really was done to not have to pay a performance penalty if
> > updating thousands of refs if no reference-transaction hook exists. E.g.
> > if doing a non-atomic push of n reference, we'd have n calls to
> > access(3P). See [1] for reference.
> > 
> > I've just did another quick benchmark without the cache, and it still
> > consistently shows a non-zero performance hit without it:
> > 
> > Test                         pks-reftx-hook-interleaving   no-cache
> > --------------------------------------------------------------------------------
> > 1400.2: update-ref           2.82(2.13+0.81)               2.86(2.19+0.78) +1.4%
> > 1400.3: update-ref --stdin   0.22(0.07+0.15)               0.22(0.07+0.15) +0.0%
> 
> I'm skeptical that those results are useful. In the first test, we're
> running update-ref 1000 times, so:
> 
>   - the cache shouldn't be helping at all, since we only have one ref to
>     update (well, I guess once for "prepare" and once for "commit", so
>     really it's saving one syscall total per process).
> 
>   - I'd expect a lot of noise because we're spending most of our time in
>     starting up the process
> 
> In the second test, we run 1000 ref operations per update-ref process.
> So we should be cutting down on our hook-lookup overhead by a factor of
> 1000. Yet it shows no improvement.
> 
> That implies you're just seeing noise. And indeed, with the patch below
> I get:
> 
> Test                         HEAD^             HEAD
> --------------------------------------------------------------------
> 1400.2: update-ref           1.93(1.57+0.42)   1.91(1.55+0.42) -1.0%
> 1400.3: update-ref --stdin   0.07(0.02+0.05)   0.07(0.02+0.05) +0.0%
> 
> Running it a second time gets me +0.5%. :)

Yeah, it's also been my take that OS-level overhead is probably going to
matter more than those access calls, and I argued such back when I
proposed the hook. So I'm perfectly happy to see this caching mechanism
go.

Should I re-post a v2 with your patch and my test?

Patrick