Message size limit seems much smaller than should be in Open MPI

Question

As I understand from this answer, there is a size limit of 2^31-1 elements for the messages that you can send/receive in Open MPI. Nevertheless, when I try to send an int array that has more than exactly 4786470 elements my code simply stops working. It does not give me any error messages or segfaults, it only gets stuck in the communication and never finishes execution.

Here is a minimal example:

#include "mpi.h"
int main(int argc, char **argv){

    MPI_Init(&argc,&argv);
    int rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);

    const int n = 4786470 + 1; // it works when n = 4786470

    int* send_array = new int[n];
    int* recv_array = new int[n];
    for(int i=0; i<n; i++) send_array[i] = 0;
    for(int i=0; i<n; i++) recv_array[i] = 0;

    if(rank==0) MPI_Send(send_array, n, MPI_INT, 1, 0, MPI_COMM_WORLD);
    if(rank==1) MPI_Recv(recv_array, n, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);

    delete[] send_array;
    delete[] recv_array;
    MPI_Finalize();
}

I did some tests with Isend and Irecv and it does not seem to change anything. I'm not sure if I am doing something wrong in the code or it's some problem from my MPI.

By the way, to compile and run I am using Open MPI 4.0.1 and Apple clang 11.0.3 in macOS 10.15.4.

score 0 · Accepted Answer · answered May 05 '20 at 23:27

It looks like this is a known issue from version 4.0.1 of Open MPI running on macOS. MPI stops communication after transmitting 19145881 bytes, which gives exactly an array of 4786470 integers. The bug was fixed for the versions >= 4.0.2 and after installing a newer version the example worked perfectly.

Message size limit seems much smaller than should be in Open MPI

1 Answers1