bug#36597: 27.0.50; rehash hash tables eagerly in pdumper

[Pip Cet <pipcet@gmail.com>]

On Sun, Jul 14, 2019 at 3:49 PM Paul Eggert <eggert@cs.ucla.edu> wrote:
> Pip Cet wrote:
> > Indeed, that's plenty of small Emacs processes not doing very much.
> > It's not the case we ought to be optimizing for, I think, but the
> > performance concerns should be taken seriously.
>
> What's a good benchmark for what we should be optimizing for? Ideally something
> somewhat-realistic as opposed to a microbenchmark.

I'd suggest something along the lines of:
perf stat -r 10 --table -e
cycles:u,instructions:u,branches:u,branch-misses:u make -B
../lisp/leim/ja-dic/ja-dic.el

With my patch:

    61,136,837,192      cycles:u
               ( +-  0.42% )
    42,313,912,525      instructions:u            #    0.69  insn per
cycle           ( +-  0.00% )
    12,131,893,779      branches:u
               ( +-  0.00% )
        47,602,747      branch-misses:u           #    0.39% of all
branches          ( +-  1.11% )

without my patch:

    61,460,927,899      cycles:u
               ( +-  0.44% )
    42,358,289,131      instructions:u            #    0.69  insn per
cycle           ( +-  0.00% )
    12,134,582,441      branches:u
               ( +-  0.00% )
        48,540,232      branch-misses:u           #    0.40% of all
branches          ( +-  1.09% )

A 0.5% improvement.

By comparison,

perf stat -r 100 --table -e
cycles:u,instructions:u,branches:u,branch-misses:u
~/git/emacs/src/emacs -Q --batch

With my patch:

        80,749,425      cycles:u
               ( +-  0.81% )
       146,770,045      instructions:u            #    1.82  insn per
cycle           ( +-  0.00% )
        29,218,226      branches:u
               ( +-  0.00% )
           450,275      branch-misses:u           #    1.54% of all
branches          ( +-  0.11% )

without my patch:

        78,896,395      cycles:u
               ( +-  0.12% )
       147,059,777      instructions:u            #    1.86  insn per
cycle           ( +-  0.00% )
        29,287,917      branches:u
               ( +-  0.00% )
           450,194      branch-misses:u           #    1.54% of all
branches          ( +-  0.09% )

About a 2% slowdown.
perf stat -r cycles:u,instructions:u,branches:u,missed-branches:u

> is dominated by a single CPU-intensive Emacs process and takes about 19 CPU
> seconds on my home desktop. The proposed patch slows this benchmark down by
> about 0.6%. (I ran the benchmark ten times after a warmup run, and took the
> average of the ten user+system times.)

Hmm. I'd like to know the reason for that, but I suspect it may simply
be thermal throttling. That's the reason I'm running tests in
parallel, though it might be better to compare instruction counts or
scheduled µ-ops rather than cycles...