使用OpenMP库，执行时间如何取决于线程数量的增加？

Question

Increasing number of threads increase the time of loop execution rather decrease it.

#include <time.h>
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#include <limits.h>
#define n 4

int main(int argc, char **argv)
{
    FILE * file1 = fopen("output.txt", "w");
    if (file1 == NULL){
        exit(EXIT_FAILURE);
    }

    srand(time(NULL));
    int matrix[n][n];
    int i, j;
    for(i = 0; i < n; i++){
        for (j = 0; j < n; j++){
            matrix[i][j] = rand() % 100 + 1;
            fprintf(file1, "%d ", matrix[i][j]);
        }
        fprintf(file1, "\n");
    }
    int sum = 0;
    int min;
    double start;
    double end;

Starting cout the time of loop

    start = omp_get_wtime();

// in num_threads I've changed the number of threads 
// and investigate a problem of increasing the time

#pragma omp parallel for private (i, j, min) reduction(+:sum)       num_threads(4) 
        for(i = 0; i < n; i++){
            min = INT_MAX;
            for (j = 0; j < n; j++){
                if(matrix[j][i] < min){
                    min = matrix[j][i];
                    }
            }
            sum += min; // sum of min numbers of each column
        }
end = omp_get_wtime();

printf("Time: %lf\n", end - start);

printf("Min sum of matrix = %d", sum);
fclose(file1);
return 0;
}

---

4 threads
Time: 0.000930
3 threads
Time: 0.000356
2 threads
Time: 0.000533
1 thread
Time: 0.000008

My CPU has 4 threads.

Answer 1

You have a very small problem (4x4) and you are timing thread creation. I don't expect the parallelism to help much at this scale anyway (since just the cost of waking the threads and then synchronizing them again at the end of the parallel will be hugely larger than the work you are trying to do), but you can remove the cost of creating the thread pool from your measurement by adding a

#pragma omp parallel ;

before the timed region.

And, please, please, please, don't force the number of threads. Use the OMP_NUM_THREADS envirable.