Question

This program estimates Pi by throwing random “darts” (sampling points) to a circle or radius=1 inscribed inside a square board of length=2. Using the relationship

Area of circle / Area of Square = Pi/4

we can estimate Pi using the same relationship expressed as

Darts Inside Circle / Darts Outside Circle = Pi/4

The program works fine when I specify NDARTS in a #define. However, when I specify NDARTS as a value that’s read via scanf and then broadcasted, it mysteriously gets stuck when more than one process is assigned via mpirun:

mpirun -np 1 ./pi_montecarlo.x

   Monte Carlo Method to estimate Pi 

Introduce Number of Darts 
10000
  Number of processes: 1 
  Number of darts: 10000 
Known value of PI  : 3.1415926535 
Estimated Value of PI  : 3.1484000000
Error Percentage   : 0.21668457
Time    : 0.00060296



mpirun -np 2 ./pi_montecarlo.x

Monte Carlo Method to estimate Pi 

Introduce Number of Darts 
10000
Number of processes: 2 
Number of darts: 10000 

^Stuck here.

Why? Is this some mpi-implementation-specific problem? Should I try another MPI implementation (I think I’m running lam)? Can you run this with at least 2 processes on your own box?

/*
mpicc -g -Wall -lm pi_montecarlo3.c -o pi_montecarlo.x 

mpirun -np 4 ./pi_montecarlo.x
*/

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <time.h>
#include <mpi.h>

#define MASTER 0
#define PI 3.1415926535

double pseudo_random (double a, double b) {
    double r; 
    r = ((b-a) * ((double) rand() / (double) RAND_MAX)) +a;
    return r; 
}

int main(int argc, char*argv[]){
    long long int NDARTS;

    int proc_id, 
        n_procs, 
        llimit,  
        ulimit,  
        n_circle, 
        i;      


    double pi_current, 
           pi_sum,     
           x,         
           y,         
           z,          
           error,      
           start_time, 
           end_time;   

    struct timeval stime;

    llimit = -1;
    ulimit = 1;
    n_circle =0; 

    MPI_Init(&argc, &argv); 

    MPI_Comm_rank (MPI_COMM_WORLD, &proc_id);
    MPI_Comm_size (MPI_COMM_WORLD, &n_procs);

    if (proc_id == MASTER){
        printf("\nMonte Carlo Method to estimate Pi \n\n");

            printf("Introduce Number of Darts \n");

            scanf("%lld",&NDARTS); 

        printf("  Number of processes: %d \n", n_procs);
        printf("  Number of darts: %lld \n", NDARTS);

            MPI_Bcast(&NDARTS, 1, MPI_LONG_LONG_INT, 0, MPI_COMM_WORLD);

            start_time = MPI_Wtime();
    }

    gettimeofday(&stime, NULL); 
    srand(stime.tv_usec * stime.tv_usec * stime.tv_usec * stime.tv_usec);

    for (i=1; i<=NDARTS;i++){
        x = pseudo_random(llimit, ulimit);
        y = pseudo_random(llimit, ulimit);

        z = pow(x,2) + pow(y,2);

        if (z<=1.0){
            n_circle++;
        }
    }

    pi_current = 4.0 * (double)n_circle / (double) NDARTS; 

    MPI_Reduce (&pi_current, &pi_sum, 1, MPI_DOUBLE, MPI_SUM, MASTER, MPI_COMM_WORLD);

       if (proc_id == MASTER) {
        pi_sum = pi_sum / n_procs;

        error = fabs ((pi_sum -PI) / PI) *100;

        end_time = MPI_Wtime();

        printf("Known value of PI  : %11.10f \n", PI);
        printf("Estimated Value of PI  : %11.10f\n", pi_sum);
        printf("Error Percentage   : %10.8f\n", error);
        printf("Time    : %10.8f\n\n", end_time - start_time);

    }

    MPI_Finalize();

    return 0;
}

Answer 1

Broadcast doesn’t “push” data onto other processors.

Almost all MPI communications requires the active participation of all processors. To send a message between two processors, for instance, the sender must call something like MPI_Send() and the receiver must call something like MPI_Recv().

This is true for collective communications, too; for instance, you have everyone calling MPI_Reduce(). Similarly, you have to have everyone call the MPI_Bcast(), not just the one that has the original data, the “receivers” too:

if (proc_id == MASTER){
    /* ... */
    scanf("%lld",&NDARTS); 
}

MPI_Bcast(&NDARTS, 1, MPI_LONG_LONG_INT, 0, MPI_COMM_WORLD);

if (proc_id == MASTER) {
    start_time = MPI_Wtime();
}

/* ... */

By the way, when you seed your random number generator, which is otherwise fine, you might want to make sure the seed is different on every processor by putting proc_id in there somewhere rather than just counting on the clocks being different enough to throw the seeds off…