Re: [PATCH v2] Add random locking benchmark

[Carlos O'Donell <carlos@redhat.com>]

On 8/22/22 11:14, Wilco Dijkstra via Libc-alpha wrote:
> v2: rename as suggested, use random throughout
> 
> Add a simple benchmark to measure the overhead of internal libc
> locks in the random() implementation on both single- and
> multi-threaded cases. This relies on the implementation of random
> using internal locks to access shared global data.

Suggest:
                         ... This relies on the implementation of random
using internal locks to access shared global data, and that the runtime
enables locking once a single thread has been created.

~~~

I think this needs a few more explanatory comments. A lot of other people
look at our benchmarks and might scratch their heads at this one.

Post a v3 and I'll ACK that.

> 
> ---
> 
> diff --git a/benchtests/Makefile b/benchtests/Makefile
> index d99771be74b40f8afa3953f61c0721b19658d4b7..c413eac1d23568cb88bf22c6e50303e24ec10ea0 100644
> --- a/benchtests/Makefile
> +++ b/benchtests/Makefile
> @@ -236,6 +236,7 @@ hash-benchset := \
>  stdlib-benchset := \
>    arc4random \
>    strtod \
> +  random-lock \

OK.

>    # stdlib-benchset
>  
>  stdio-common-benchset := sprintf
> diff --git a/benchtests/bench-random-lock.c b/benchtests/bench-random-lock.c
> new file mode 100644
> index 0000000000000000000000000000000000000000..13c8e77cbd3538bb1a15f81def0a2c41a3763729
> --- /dev/null
> +++ b/benchtests/bench-random-lock.c
> @@ -0,0 +1,102 @@
> +/* Benchmark internal libc locking functions used in random.

OK.

> +   Copyright (C) 2022 Free Software Foundation, Inc.> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#define TEST_MAIN
> +#define TEST_NAME "random-lock"
> +
> +#include <pthread.h>
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include "bench-timing.h"
> +#include "json-lib.h"
> +

We could run the benchmark for a fixed amount of time, rather than a
fixed number of iterations, see bench-skeleton.c and DURATION.

Suggest:

/* Arbitrarily pick 20M such that we can measure the time taken.  */

> +#define NUM_ITERS 20000000

OK.

> +
> +json_ctx_t json_ctx;
> +
> +
> +/* Measure the overhead of __libc_lock_lock and __libc_lock_unlock by
> +   calling random ().  */
> +static void
> +bench_random_lock (size_t iters)
> +{
> +  timing_t start, stop, total;
> +
> +  srandom (0);

OK. Seed to 0.


Suggest:

/* Warmup. Arbitrarily pick 1/4 of iterations.  */

> +  for (int i = 0; i < iters / 4; i++)
> +    (void) random ();



> +
> +  TIMING_NOW (start);
> +
> +  for (int i = 0; i < iters; i++)
> +    (void) random ();
> +
> +  TIMING_NOW (stop);
> +
> +  TIMING_DIFF (total, start, stop);
> +
> +  json_element_double (&json_ctx, (double) total / (double) iters);
> +}
> +
> +static void *
> +thread_start (void *p)
> +{
> +  return p;
> +}
> +
> +int
> +do_bench (void)
> +{
> +  json_init (&json_ctx, 0, stdout);
> +
> +  json_document_begin (&json_ctx);
> +
> +  json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
> +  json_attr_object_begin (&json_ctx, "functions");
> +  json_attr_object_begin (&json_ctx, "random");
> +  json_attr_string (&json_ctx, "bench-variant", "single-threaded");
> +  json_array_begin (&json_ctx, "results");
> +
> +  /* Run benchmark single threaded.  */
> +  bench_random_lock (NUM_ITERS);

OK.

> +
> +  json_array_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +
> +  json_attr_object_begin (&json_ctx, "random");
> +  json_attr_string (&json_ctx, "bench-variant", "multi-threaded");
> +  json_array_begin (&json_ctx, "results");
> +
> +  pthread_t t;

Suggest:

/* We assume the runtimes decision to keep the library multi-threaded
   once a single thread has launched, and not change back to
   single-threaded.  */

Provides a little more context to this section, even though
you mention 'SINGLE_THREAD_P == false' later.

> +  pthread_create (&t, NULL, thread_start, NULL);
> +  pthread_join (t, NULL);

OK. We are now in multi-threaded and will never switch back to the
single threaded case.

> +
> +  /* Repeat benchmark now SINGLE_THREAD_P == false.  */
> +  bench_random_lock (NUM_ITERS);
> +
> +  json_array_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_document_end (&json_ctx);
> +  return 0;
> +}
> +
> +#define TEST_FUNCTION do_bench ()
> +
> +#include "../test-skeleton.c"
> +
> 
> 
> 
> 


-- 
Cheers,
Carlos.