Re: [PATCH] grep: use no. of cores as the default no. of threads

[Matheus Tavares <matheus.bernardino@usp.br>]

Hi, Victor

On Thu, Jan 16, 2020 at 10:11 AM Victor Leschuk <vleschuk@gmail.com> wrote:
>
> Grepping bottleneck is not cpu, but IO. Maybe it is more reasonable to
> use not online_cpus() but online_cpus()*2?

I also tried this approach, but the tests I ran with online_cpus() * 2
only showed slowdowns. The results can be seen in the commit message:

> On Thu, Jan 16, 2020 at 5:41 AM Matheus Tavares
> <matheus.bernardino@usp.br> wrote:
> >
[...]
> > The following measurements correspond to the
> > mean elapsed times for 30 git-grep executions in chromium's
> > repository[1] with a 95% confidence interval (each set of 30 were
> > performed after 2 warmup runs). Regex 1 is 'abcd[02]' and Regex 2 is
> > '(static|extern) (int|double) \*'.
> >
> >       |          Working tree         |           Object Store
> > ------|-------------------------------|--------------------------------
> >  #ths |  Regex 1      |  Regex 2      |   Regex 1      |   Regex 2
> > ------|---------------|---------------|----------------|---------------
> >   32  |  2.92s ± 0.01 |  3.72s ± 0.21 |   5.36s ± 0.01 |   6.07s ± 0.01
> >   16  |  2.84s ± 0.01 |  3.57s ± 0.21 |   5.05s ± 0.01 |   5.71s ± 0.01
> > >  8  |  2.53s ± 0.00 |  3.24s ± 0.21 |   4.86s ± 0.01 |   5.48s ± 0.01
> >    4  |  2.43s ± 0.02 |  3.22s ± 0.20 |   5.22s ± 0.02 |   6.03s ± 0.02
> >    2  |  3.06s ± 0.20 |  4.52s ± 0.01 |   7.52s ± 0.01 |   9.06s ± 0.01
> >    1  |  6.16s ± 0.01 |  9.25s ± 0.02 |  14.10s ± 0.01 |  17.22s ± 0.01
> >
> > The above tests were performed in a desktop running Debian 10.0 with
> > Intel(R) Xeon(R) CPU E3-1230 V2 (4 cores w/ hyper-threading), 32GB of
> > RAM and a 7200 rpm, SATA 3.1 HDD.
> >
> > Bellow, the tests were repeated for a machine with SSD: a Manjaro laptop
> > with Intel(R) i7-7700HQ (4 cores w/ hyper-threading) and 16GB of RAM:
> >
> >       |          Working tree          |           Object Store
> > ------|--------------------------------|--------------------------------
> >  #ths |  Regex 1      |  Regex 2       |   Regex 1      |   Regex 2
> > ------|---------------|----------------|----------------|---------------
> >   32  |  3.29s ± 0.21 |   4.30s ± 0.01 |   6.30s ± 0.01 |   7.30s ± 0.02
> >   16  |  3.19s ± 0.20 |   4.14s ± 0.02 |   5.91s ± 0.01 |   6.83s ± 0.01
> > >  8  |  2.90s ± 0.04 |   3.82s ± 0.20 |   5.70s ± 0.02 |   6.53s ± 0.01
> >    4  |  2.84s ± 0.02 |   3.77s ± 0.20 |   6.19s ± 0.02 |   7.18s ± 0.02
> >    2  |  3.73s ± 0.21 |   5.57s ± 0.02 |   9.28s ± 0.01 |  11.22s ± 0.01
> >    1  |  7.48s ± 0.02 |  11.36s ± 0.03 |  17.75s ± 0.01 |  21.87s ± 0.08

I deliberately used somehow complex regexes for these tests. So I
decided to do one more test with a very simple fixed string ("abc"),
allowing git-grep to spend less time in the cpu-bound regex searching.
The results can be seen bellow (the metodology is the same as described
above and the machine is the Manjaro laptop, for which online_cpus()
returns 8):

  #ths |  Working Three |  Object Store
 ------|----------------|---------------
    16 |  3.22s ± 0.20  |  5.96s ± 0.06
     8 |  2.92s ± 0.01  |  5.73s ± 0.02