Topology

C | Fortran-2008 | Fortran-90

MPI_Cart_sub

Definition

MPI_Cart_sub allows to partition a cartesian topology created with MPI_Cart_create. It works by selecting dimensions along which the subgrids created will be arranged. For each subgrid, it creates a group and a communicator.

Copy

Feedback

int MPI_Cart_sub(MPI_Comm comm,
                 const int remain_dims[],
                 MPI_Comm* new_comm);

Parameters

comm

The cartesian communicator, created with MPI_Cart_create, to partition.

remain_dims

Indicates along which dimensions arrange subgrids. If all entries in remain_dims are set to false, then new_comm will be associated with a zero-dimensional cartesian topology.

new_comm

The communicator containing the subgrid that includes the calling MPI process.

Return value

The error code returned from the cartesian partitioning.

Example

Copy

Feedback

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

/**
 * @brief Illustrates how to partition a cartesian topology created with
 * MPI_Cart_create.
 * @details This program is meant to be run with 6 MPI processes. It consists in
 * partitioning a 2 x 3 2D cartesian topology along its first dimension to
 * obtain 2 1D subgrids of 3 MPI processes each.
 * The initial 2D cartesian topology can be visualised as follows:
 *
 *                  Second dimension
 * ------------------------------------------------->
 *
 * +---------------+---------------+---------------+   |
 * | MPI process 0 | MPI process 1 | MPI process 2 |   |
 * |     (0,0)     |     (0,1)     |     (0,2)     |   |
 * +---------------+---------------+---------------+   |  First dimension
 * | MPI process 3 | MPI process 4 | MPI process 5 |   |
 * |     (1,0)     |     (1,1)     |     (1,2)     |   |
 * +---------------+---------------+---------------+   v
 *
 * And the final subgrids can be visualised as follows:
 *
 * +---------------+---------------+---------------+  \
 * | MPI process 0 | MPI process 1 | MPI process 2 |   } First subgrid
 * +---------------+---------------+---------------+  /
 *
 * +---------------+---------------+---------------+  \
 * | MPI process 3 | MPI process 4 | MPI process 5 |   } Second subgrid
 * +---------------+---------------+---------------+  /
 *
 **/
int main(int argc, char* argv[])
{
    MPI_Init(&argc, &argv);

    // Size of the default communicator
    int size;
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    if(size != 6)
    {
        printf("This application is meant to be run with 6 processes.\n");
        MPI_Abort(MPI_COMM_WORLD, EXIT_FAILURE);
    }

    // Ask MPI to decompose our processes in a 2D cartesian grid for us
    int dims[2] = {2, 3};

    // Make both dimensions non-periodic
    int periods[2] = {false, false};

    // Let MPI assign arbitrary ranks if it deems it necessary
    int reorder = true;

    // Create a communicator given the 2D torus topology.
    MPI_Comm cartesian_communicator;
    MPI_Cart_create(MPI_COMM_WORLD, 2, dims, periods, reorder, &cartesian_communicator);

    // My rank in the new communicator
    int my_rank;
    MPI_Comm_rank(cartesian_communicator, &my_rank);

    // Get my coordinates in the new communicator
    int my_coords[2];
    MPI_Cart_coords(cartesian_communicator, my_rank, 2, my_coords);

    // Print my location in the 2D cartesian topology.
    printf("[MPI process %d] I am located at (%d, %d) in the initial 2D cartesian topology.\n", my_rank, my_coords[0], my_coords[1]);

    // Partition the 2D cartesian topology along the first dimension, by preserving the second dimension
    int remain_dims[2] = {false, true};
    MPI_Comm subgrid_communicator;
    MPI_Cart_sub(cartesian_communicator, remain_dims, &subgrid_communicator);

    // Get the ranks of all MPI processes in my subgrid and print it
    int subgrid_ranks[3];
    MPI_Allgather(&my_rank, 1, MPI_INT, subgrid_ranks, 1, MPI_INT, subgrid_communicator);
    printf("[MPI process %d] I am in the 1D subgrid that contains MPI processes %d, %d and %d.\n", my_rank, subgrid_ranks[0], subgrid_ranks[1], subgrid_ranks[2]);

    MPI_Finalize();

    return EXIT_SUCCESS;
}