Collectives

Non-blocking

Synchronisation

C | Fortran-2008 | Fortran-90

MPI_Ibarrier

Definition

MPI_Ibarrier is the non-blocking version of MPI_Barrier; it does not wait for all other processes to call MPI_Ibarrier to return. A non-blocking barrier may appear counter-productive at first, but it allows to overlap the barrier synchronisation with independent computation.

Copy

Feedback

int MPI_Ibarrier(MPI_Comm communicator,
                 MPI_Request* request);

Parameters

communicator

The communicator in which the barrier takes place.

request

The variable in which store the non-blocking operation handle.

Return value

The error code returned from the non-blocking barrier.

Example

Copy

Feedback

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

/**
 * @brief Illustrates how to use a non-blocking barrier.
 * @details This application provides a basic example about a non-blocking
 * barrier. Although this example may not be very realistic, it does illustrate
 * how to manipulate a non-blocking barrier.
 *
 * This application consists of three jobs for each process:
 * 1) Get my rank
 * 2) When I have my rank, I print it
 * 3) Once all processes obtained their rank, I say it
 *
 * Rather than using a classic barrier after 2), we can issue a non-blocking
 * barrier after 1) so that we overlap the synchronisation of the barrier with
 * the task 2). Then, once 2) is complete, we can block until the non-blocking
 * barrier completes to progress to 3).
 *
 * The difference can be visualised as follows:
 *
 *               +--------------+
 *               | Barrier sync |
 * +------+------+--------------+------+
 * | Task | Task |              | Task |
 * |   1  |   2  |              |   3  |
 * +------+------+--------------+------+
 *                                     |
 *        +--------------+             |
 *        | Barrier sync |             |
 * +------+------+-------+------+      |
 * | Task | Task |       | Task |      |
 * |  1   |   2  |       |   3  |      |
 * +------+------+--------------+      |
 *                              |      |
 *                              V      V
 * ---------------- TIME -------+------+----->
 *                               <---->
 *                                gain
 **/
int main(int argc, char* argv[])
{
    MPI_Init(&argc, &argv);

    // Get my rank, this is task 1
    int my_rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
    MPI_Request request;
    MPI_Ibarrier(MPI_COMM_WORLD, &request);

    // Task 2
    printf("[MPI process %d] I got my rank, it is %d, I now call MPI_Ibarrier.\n", my_rank, my_rank);

    // Task 3
    MPI_Wait(&request, MPI_STATUS_IGNORE);
    printf("[MPI process %d] The MPI_Ibarrier is complete; all processes got their rank.\n", my_rank);

    MPI_Finalize();

    return EXIT_SUCCESS;
}