Re: [PATCH v2 2/5] benchtests: Add memset zero fill benchtest

[Noah Goldstein via Libc-alpha <libc-alpha@sourceware.org>]

On Wed, Jul 21, 2021 at 9:07 AM naohirot@fujitsu.com <naohirot@fujitsu.com>
wrote:

> Hi Noah,
>
> One typo in the updated code.
> Wrong:
>   #define START_SIZE (16 * 1024)
> Right:
>   #define BUF1PAGES 16
>
> Thanks
> Naohiro
> ________________________________________
> From: Tamura, Naohiro/田村 直広 <naohirot@fujitsu.com>
> Sent: Wednesday, 21 July 2021 21:56
> To: Noah Goldstein
> Cc: Wilco Dijkstra; Lucas A. M. Magalhaes; GNU C Library
> Subject: RE: [PATCH v2 2/5] benchtests: Add memset zero fill benchtest
>
> Hi Noah,
>
> Thank you for the review.
>
> > > +#define TEST_MAIN
> > > +#define TEST_NAME "memset"
> > > +#define START_SIZE (16 * 1024)
> > > +#define MIN_PAGE_SIZE (getpagesize () + 64 * 1024 * 1024)
> > > +#define TIMEOUT (20 * 60)
> > > +#include "bench-string.h"
> > > +
> > > +#include "json-lib.h"
> > > +
> > > +void *generic_memset (void *, int, size_t);
> > > +typedef void *(*proto_t) (void *, int, size_t);
> > > +
> > > +IMPL (MEMSET, 1)
> > > +IMPL (generic_memset, 0)
> > > +
> > > +static void
> > Do we want __attribute__((noinline, noclone))?
>
> Yes, I'll add it.
>
> > > +do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
> > > +            int c1 __attribute ((unused)), int c2 __attribute
> ((unused)),
> > > +            size_t n)
> > > +{
> > > +  size_t i, iters = 16;
> >
> > I think 16 is probably too few iterations for reliable benchmarking.
> > Maybe `INNER_LOOP_ITERS` which is 8192
>
> I tried it. If it is changed to 8192, it hit the TIMEOUT (20 * 60) on
> a64fx.
> Please check the code below.
>
> >
> > > +  timing_t start, stop, cur;
> > > +
> > > +  TIMING_NOW (start);
> > > +  for (i = 0; i < iters; i += 2)
> > > +    {
> > > +      CALL (impl, s, c1, n);
> > I am a bit worried that the overhead from the first call with `c1` will
> distort the results.
> > Is it possible to implement it with a nested loop where you fill `s`
> with `c1` for
> > `n * inner_loop_iterations` in the outer loop and in the inner loop fill
> `c2` on `s + n * i`?
> > In that case maybe 16 for inner loop iterations and 512 for outer loop
> iterations.
>
> It seems that we have to set smaller number if this implementation is not
> wrong.
> Because it will take 99.4 minutes estimating from the case that "iters =
> 32"
> took 23.3 seconds.
> (8192/32*23.3/60=99.4)

I see. I think 16 for the inner loop makes sense. From the x86_64
perspective this
will keep the loop from running out of the LSD which is necessary for
accurate
benchmarking. I guess then somewhere between [2, 8] is reasonable for the
outer
loop?


> #define START_SIZE (16 * 1024)
> ...
> static void
> __attribute__((noinline, noclone))
> do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
>              int c1 __attribute ((unused)), int c2 __attribute ((unused)),
>              size_t n)
> {
>   size_t i, j, iters = INNER_LOOP_ITERS; // 32;
>   timing_t start, stop, cur, latency = 0;
>
>   for (i = 0; i < 512; i++) // for (i = 0; i < 2; i++)
>     {

      CALL (impl, s, c1, n * 16);
>       TIMING_NOW (start);
>       for (j = 0; j < 16; j++)
>         CALL (impl, s + n * j, c2, n);
>       TIMING_NOW (stop);
>       TIMING_DIFF (cur, start, stop);
>       TIMING_ACCUM (latency, cur);
>     }
>
This looks good. But as you said, a much smaller value for outer loop.

>
>   json_element_double (json_ctx, (double) latency / (double) iters);
> }
>
> > > +      CALL (impl, s, c2, n);
> > > +    }
> > > +  TIMING_NOW (stop);
> > > +
> > > +  TIMING_DIFF (cur, start, stop);
> > > +
> > > +  json_element_double (json_ctx, (double) cur / (double) iters);
> > > +}
> > > +
> > > +static void
> > > +do_test (json_ctx_t *json_ctx, size_t align, int c1, int c2, size_t
> len)
> > > +{
> > > +  align &= 63;
> > Can you make this `align &= getpagesize () - 1;`?
>
> I'll change it.
>
> Thanks.
> Naohiro
>