CODEWITHSUNDEEP
cunixparallel-processingmpiopenmpi
Jack Hoff.
Jack Hoff.

Reputation: 143

Understanding buffer passing via MPI_Isend and MPI_Irecv

I want to understand how sending and receiving are processed by MPI. Suppose I allocate a buffer of [12][50] elements as the following:

int **buf= malloc(12 * sizeof(int *));
for (i = 0; i < 12; i++)
{
    buf[i] = malloc(50 * sizeof(int));

    // Immediatly fill each row by 1s for testing purpose. 
    for (j = 0; j < 50; j++)
    {
         buf[i][j] = 1;
    }
}

Now, I want to send each row to P = 12 processors using non-blocking MPI_Isend and MPI_Irecv as the following:

for (i = 0; i < 12; i++)
{
    MPI_Isend(buf[i], 50, MPI_INT, i, TAG_0, MPI_COMM_WORLD, &req[i]); 
    MPI_Wait(&req[i], &status[i]);
}

    for (i = 0; i < 12; i++)
    {
        MPI_Irecv(bufRecv[i], 50, MPI_INT, MASTER, TAG_0, MPI_COMM_WORLD, &req[i]);
        MPI_Wait(&req[i], &status[i]);
    }

As far as I know, MPI_Isend here sends each ith row followed by 50 consecutive elements starting from the memory address stored in but[i] whereas MPI_Irecv receives the corresponding ith row and store it in MPI_Irecv . Am I right? If no, can someone explain why?

Thank you.

Upvotes: 0

Views: 918

Answers (1)

haraldkl
haraldkl

Reputation: 3819

Yes, you are right. MPI_Irecv would receive the message into bufRecv. However, your code does not really make sense. First off, a non-blocking communication call directly followed by a wait is basically the same as a blocking call.

Then the receiving part should be called on a different process, and from your description each receiver would call MPI_Irecv only once, not 12 times. Or, if every process is to receive all rows, the sender has to call send for each process number of rows times.

Most importantly: what you describe is covered by MPI_Scatter, no need to implement it yourself. If you want one process to send data to all processes, use MPI_Bcast.

Upvotes: 0

Related Questions