Learn HPCToolkit

[ian-bertolacci]

From April first email:

• best possible parallel implementation with OpenMP
  • investigate with hpctoolkit when synchronization is an issue with the larger problem sizes

Problem: I don't know hpctoolkit or how to use it to identify synchronization issues.

Done when:

• I can show when synchronization becomes an issue on the following OpenMP programs which certainly has some synchronization overhead.

Critical section

#include <omp.h>
#include <stdio.h>
#include <stdlib.h>

volatile int blah;

int main(){
  int T = 1000;
  int N = 10000;

  int* array = (int*) calloc(sizeof(int), N);

  for( int t = 0; t < T; ++t ){

    #pragma omp parallel for
    for( int i = 0; i < N; ++i ){
      #pragma omp critical
      blah += array[i];
    }

  }
}

Reduction

#include <omp.h>
#include <stdio.h>
#include <stdlib.h>

volatile int blah;

int main(){
  int T = 1000;
  int N = 10000;

  int* array = (int*) calloc(sizeof(int), N);

  for( int t = 0; t < T; ++t ){

    #pragma omp parallel for reduction(+:blah)
    for( int i = 0; i < N; ++i ){
      blah += array[i];
    }

  }
}