Re: [PATCH v2 4/4] Bump rename limit defaults (yet again)

From: Elijah Newren <newren@gmail.com>
Date: Jul 14, 2021, 20:03 UTC
Message-ID: <CABPp-BE=Ghkm=sZ05=A=1vC1zscbVk7qXgeEjpA5gHxsfTGAyg@mail.gmail.com>
In-Reply-To: <YO8lgYa5/RlL1zGp@coredump.intra.peff.net>

On Wed, Jul 14, 2021 at 10:57 AM Jeff King <peff@peff.net> wrote:

Show 31 quoted lines

>
> On Wed, Jul 14, 2021 at 10:32:56AM -0700, Elijah Newren wrote:
>
> > > It's slightly sad that we only got a 30% CPU improvement in the past 10
> > > years. Are you just counting clock speed as a short-hand here? I think
> > > that doesn't tell the whole story. But all of this is a side-note
> > > anyway.  What I care about is your actual timings. :)
> >
> > I'm using shorthand when discussing file sizes above (though I used
> > actual measurements when picking new values below).  But the 30% came
> > from measuring the timings with the exact same sample file as you and
> > using a lightly modified version of your original script (tweaked to
> > also change file basenames) on an AWS c5xlarge instance.  My timings
> > showed they were only about 30% faster than what you got when you last
> > bumped the limits.
>
> Interesting. My timings are much faster. With a 20k file, I get (on my
> laptop, which is an i9-9880H):
>
>      N     CPU (2008)    CPU (now)
>     10          0.43s       0.007s
>    100          0.44s       0.071s
>    200          1.40s       0.226s
>    400          4.87s       0.788s
>    800         18.08s       2.887s
>   1000         27.82s       4.431s
>
> The 2008 timings are from the old email you linked in your commit
> message, and the new one is from running the revised script you showed.
> The savings seem like more than 30%. I don't know if that's all CPU or
> if something changed in the code.

I was using the script you wrote in 2008, but comparing to your reported numbers in 2011[1]. When you bumped in 2011, you said you picked the limits of 400 & 1000 in order to give rough timings of 2s and 10s. So the table looks more like:

   N  CPU (2008)  CPU (2011)  c5xlarge  CPU (yours)
 400    4.87s         ~2s      1.106s      0.788s
1000   27.82s        ~10s      6.350s      4.431s

So, 2011->c5xlarge on these (just recomputed now numbers) show improvements of ~45% and ~36%. Maybe I had an outlier run earlier that was in the upper 6s range for N=1000 and I rounded off to 30% for the commit message? Don't know, those numbers are on a laptop that died in the last few days.

But yeah, you are seeing a bigger improvement than I did; 2011->your current laptop shows roughly 56% - 60% improvement.

[1] https://lore.kernel.org/git/20110219102128.GB22508@sigill.intra.peff.net/

> Using a 3k file (the median for ls-tree), numbers are similar, but a
> little smaller (my n=1300 is about 1.4s).  So I think we're both in the
> same ballpark (and certainly an AWS machine is a perfectly fine
> representative sample of where people might run Git).

Yeah, dividing any of the timings I get by the ones you get seem to be giving a value somewhere around 1.4, so that seems reassuring on the consistency front. That doesn't fix my jealousy of your faster CPU, but that's a separate problem. :-)