openmp的运行时作为线程数的函数正在增加

Question

我有一个运行循环的c程序并计时它的运行时间。要查看加速，我使用omp_set_num_threads更改主体中的线程数，然后调用函数。

int main(){
      ...
      for(int num_threads = 1; num_threads<=16; num_threads *= 2)
      {
        omp_set_num_threads(num_threads);
        pFilterFirst ( ..., num_threads );
        checkData ( serial_array, output_array );
        memset ( output_array, 0, DATA_LEN );

        pDataFirst ( ..., num_threads );
        checkData ( serial_array, output_array );
        memset ( output_array, 0, DATA_LEN );
      }
    }



int pDataFirst(..., num_threads) {
    /* Variables for timing */
  struct timeval ta, tb, tresult;

  /* get initial time */
  gettimeofday ( &ta, NULL );

  #pragma omp parallel for

  /* for all elements in the data */
  for (int x=0; x<data_len; x++) {
    /* for all elements in the filter */
    for (int y=0; y<filter_len; y++) {
      /* it the data element matches the filter */
      if (input_array[x] == filter_list[y]) {
        /* include it in the output */
        output_array[x] = input_array[x];
      }
    }
  }


  /* get initial time */
  gettimeofday ( &tb, NULL );

  timeval_subtract ( &tresult, &tb, &ta );

  printf ("Parallel data first took %lu seconds and %lu microseconds.  Filter l\
ength = %d. Threads = %d\n", tresult.tv_sec, tresult.tv_usec, filter_len, num_t\
hreads );
}

当我将运行时视为线程数的函数时，运行时随着线程数的增加而增加，这没有任何意义。我在亚马逊ec2实例上运行它。

因此，当我在大型ec2实例上运行模拟时，问题似乎已得到解决。我获得了预期的加速。