Re: RFC: Lightweight synchronization mechanism for gnumach

[Agustina Arzille]

On 02/28/2016 06:32 PM, Samuel Thibault wrote:
> Hello,
>
> Agustina Arzille, on Sun 28 Feb 2016 14:22:03 -0300, wrote:
> > Here's a rough comparison of how locking is implemented in libpthread,
> > and the futex prototype I posted.
> Did you make measurements? It'd be interesting to post them here.
>
> Samuel

Hello, Samuel.

Here's the benchmarking program I used to compare pthread mutexes and 
futexes:

=================================

#include <stdio.h>
#include <pthread.h>

#ifdef USE_FUTEX
  #include "futex.h"
#  define lock_t   unsigned int
#  define LOCK_INITIALIZER   0
#  define lock_acq(Lp)   futex_lock ((Lp), 0)
#  define lock_rel(Lp)   futex_unlock ((Lp), 0)
#else
#  define lock_t   pthread_mutex_t
#  define LOCK_INITIALIZER   PTHREAD_MUTEX_INITIALIZER
#  define lock_acq(Lp)   pthread_mutex_lock (Lp)
#  define lock_rel(Lp)   pthread_mutex_unlock (Lp)
#endif

static lock_t lock = LOCK_INITIALIZER;
static volatile int value = 0;

#define LIMIT   1000000

void* thrfct (void *argp)
{
  int i, off = argp ? 1 : -1;
  for (i = 0; i < LIMIT; ++i)
    {
      lock_acq (&lock);
      value += off;
      lock_rel (&lock);
    }

  return (0);
}

#define NTHREADS   5

int main (void)
{
  pthread_t thrs[NTHREADS];
  int i;

  for (i = 0; i < NTHREADS; ++i)
    pthread_create (&thrs[i], NULL, thrfct, (void *)(i & 1));

  thrfct ((void *)(NTHREADS & 1));
  for (i = 0; i < NTHREADS; ++i)
    pthread_join (thrs[i], NULL);

  printf ("RESULT: %d\n", value);
  return (0);
}

=================================

The file "futex.h" simply contains the wrappers as posted in:
http://lists.gnu.org/archive/html/bug-hurd/2016-02/msg00138.html

To test this, I compiled 2 versions of the program: One that uses 
pthread and
the other that uses futexes:

gcc -O2 -DUSE_FUTEX benchmark.c -o futexes -lpthread
gcc -O2 benchmark.c -o pthreads -lpthread

Here are the times I got from running "time ./pthreads"

I'm using gcc 5.2.1 and running Hurd with KVM.

real: 0m1.340s
user: 0
sys : 0

real: 0m3.030s
user: 0
sys : 0

real: 0m0.390s
user: 0
sys : 0

real: 0m0.380s
user: 0
sys : 0

real: 0m3.010s
user: 0
sys : 0

And here are the times for "time ./futexes"

real: 0m0.160s
user: 0
sys : 0

real: 0m0.160s
user: 0
sys : 0

real: 0m0.270s
user: 0
sys : 0

real: 0m0.150s
user: 0
sys : 0

real: 0m0.150s
user: 0
sys : 0

Now, the benchmark I used isn't exactly representative of real-life
programs (At least, I hope it isn't), but I think it serves to illustrate
the performance of each implementation under heavy contention.