CODEWITHSUNDEEP
crandomopenmp
Adrien Roussel
Adrien Roussel

Reputation: 49

Random values generator bug with OpenMP tasks

I'm trying to develop a micro code based on the calculation of Pi by the Monte Carlo method while using OpenMP tasks. It seems that I have some troubles in that code. Indeed, my application simply crashes and I got a segmentation fault coming from the gomp_barrier function...

I don't know if I am doing wrong things in my code or not, but I spend a little bit time on time to try to debug it and I found nothing. When I simply remove the computation of x and y with random values generator, the code is doing well... But if I add a call to rand_r or srand48_r, I have this segfault. So maybe, my use of these functions is wrong. Does anyone have an idea please ?

Here is my code :

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <inttypes.h>
#include <time.h>
#include <unistd.h>

#include "omp.h"

#define TRIALS_PER_THREAD 10E6

int main(int argc, char** argv)
{
  uint64_t const n_test = TRIALS_PER_THREAD;
  uint64_t i;
  double pi = 0.;

  int nb_threads = 2;
  #pragma omp parallel shared(nb_threads)
  {
    #pragma omp master
      nb_threads = omp_get_num_threads();

  }
  fprintf(stdout, "Nb threads: %d\n", nb_threads);

  uint64_t* result = (uint64_t*)malloc(sizeof(uint64_t) * nb_threads);
  for(i = 0; i < nb_threads; ++i)
  {
    result[i] = 0;
  }

  int nb_test_per_thread = 20;

  #pragma omp parallel\
  shared(result)\
  firstprivate(n_test,nb_test_per_thread)
  {
    unsigned int seed;
    struct drand48_data randBuffer;

    seed = time(NULL) ^ omp_get_thread_num() ^ getpid();
    srand48_r(seed, &randBuffer);

    for(int k = 0; k < nb_test_per_thread; ++k)
    {
      #pragma omp task shared(result, seed, randBuffer)
      {
        uint64_t local_res = 0;
        double x = 0., y = 0.;
        for(i = 0; i < n_test; ++i)
        {
          drand48_r(&randBuffer, &x);// / (double)RAND_MAX;
          drand48_r(&randBuffer, &y);// / (double)RAND_MAX;

          local_res += (((x * x) + (y * y)) <= 1);
        }
        int tid = omp_get_thread_num();
        result[tid] += local_res;
      }
    }
  }

  for(i = 0; i < nb_threads; ++i)
  {
    pi += result[i];
  }

  fprintf(stdout, "%ld of %ld throws are in the circle !\n", (uint64_t)pi, n_test*nb_test_per_thread*nb_threads);
  pi *= 4;
  pi /= (double)(n_test*nb_test_per_thread*nb_threads);
  fprintf(stdout, "Pi ~= %f\n", pi);

  return 0;
}

Thank you in advance for your help !

Upvotes: 0

Views: 319

Answers (1)

Andrew Henle
Andrew Henle

Reputation: 1

Your use of drand48_r(&randBuffer, &x); has a race condition because each OpenMP thread shares randBuffer.

Per the Linux drand48_r() man page:

SYNOPSIS

   #include <stdlib.h>

   int drand48_r(struct drand48_data *buffer, double *result);

  ...

...

ATTRIBUTES

For an explanation of the terms used in this section, see attributes(7).

   ┌──────────────────────────┬───────────────┬─────────────────────┐
   │Interface                 │ Attribute     │ Value               │
   ├──────────────────────────┼───────────────┼─────────────────────┤
   │drand48_r(), erand48_r(), │ Thread safety │ MT-Safe race:buffer │
   │lrand48_r(), nrand48_r(), │               │                     │
   │mrand48_r(), jrand48_r(), │               │                     │
   │srand48_r(), seed48_r(),  │               │                     │
   │lcong48_r()               │               │                     │
   └──────────────────────────┴───────────────┴─────────────────────┘

Note the MT-Safe race:buffer in the ATTRIUBTES table.

Per the attributes(7) man page:

   `:identifier`

Annotations may sometimes be followed by identifiers, intended to group several functions that, for example, access the data structures in an unsafe way, as in race and const ...

Multiple parallel calls to drand48_r() can not safely share the same buffer argument due to a data race condition.

Upvotes: 1

Related Questions