1. Goal

I have to distribute an array, called A_loc, over a custom communicator (that is not MPI_COMM_WORLD). Let's suppose we want to distribute an array over mesh_r communicator:

P0-P1
|  |  
P2-P3

where - represents mesh_r (mesh_rows) communicator and | represents mesh_c (mesh_columns) communicator, build through build_mesh procedure.

2. Code

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

bool is_divisible(int, int);
void build_mesh(MPI_Comm*, MPI_Comm*, MPI_Comm*, int, int, int, int, int*);
int *fill_matrix(int*, int, int);
void print_matrix(int*, int, int, int, int);
void handle_errors(int, int, int, int);
void distribute(int*, int*, int, int, int, int, int, int, int);
void debug(int*, int*, int, int, int, int, int, int, int);

int main(int argc, char *argv[])
{
    int process_rank, world_size;
    int mesh_rows, mesh_columns;
    int mesh_dimension = 2;
    int *process_coordinates;
    MPI_Comm mesh, mesh_r, mesh_c;
    int process_rank_mesh;
    int *A, *A_loc;
    int *B, *B_loc;
    int m, n, mloc, nloc;

    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &process_rank);
    MPI_Comm_size(MPI_COMM_WORLD, &world_size);

    if (process_rank == 0) {
        m = n = world_size * 1; // multiple of world_size = 4
    }

    MPI_Bcast(&m, 1, MPI_INT, 0, MPI_COMM_WORLD);
    MPI_Bcast(&n, 1, MPI_INT, 0, MPI_COMM_WORLD);
    A = fill_matrix(A, m, n);
    B = fill_matrix(A, m, n);

    if (process_rank == 0) 
        mesh_rows = 2;

    if (is_divisible(world_size, mesh_rows))
        mesh_columns = world_size / mesh_rows;
    else {
        mesh_rows = 1;
        mesh_columns = world_size / mesh_rows;
    }
   
    MPI_Bcast(&mesh_rows, 1, MPI_INT, 0, MPI_COMM_WORLD);
    MPI_Bcast(&mesh_columns, 1, MPI_INT, 0, MPI_COMM_WORLD);

    process_coordinates = (int*) calloc(mesh_dimension, sizeof(int));
    build_mesh(&mesh, &mesh_r, &mesh_c, process_rank, world_size, mesh_rows, mesh_columns, process_coordinates);
    MPI_Comm_rank(mesh, &process_rank_mesh); 
 
    mloc = m / mesh_rows;
    nloc = m / mesh_columns;

    handle_errors(m, n, world_size, process_rank);

    A_loc = (int*) calloc(mloc * nloc, sizeof(int));
    distribute(A, A_loc, m, n, mloc, nloc, world_size, mesh_rows, mesh_columns);
    
    B_loc = (int*) calloc(mloc * nloc, sizeof(int));
    distribute(B, B_loc, m, n, mloc, nloc, world_size, mesh_rows, mesh_columns);

    // I want to re-write this part so I can exploit mesh_r communicator instead of MPI_COMM_WORLD...
    int *A_loc_add = (int*) calloc(mloc * nloc, sizeof(int));
    if (process_rank == 0) {
        MPI_Send(A_loc, mloc * nloc, MPI_INT, 1, 10, MPI_COMM_WORLD);
    } else if (process_rank == 3) {
        MPI_Send(A_loc, mloc * nloc, MPI_INT, 2, 20, MPI_COMM_WORLD);
    }
    MPI_Status status;
    if (process_rank == 1) {
        MPI_Recv(A_loc_add, mloc * nloc, MPI_INT, 0, 10, MPI_COMM_WORLD, &status);
    } else if (process_rank == 2) {
        MPI_Recv(A_loc_add, mloc * nloc, MPI_INT, 3, 20, MPI_COMM_WORLD, &status);
    }

    MPI_Finalize();
    return 0;
}


void distribute(int *Mat, int *Mat_loc, int m, int n, int mloc, int nloc, int world_size, int mesh_rows, int mesh_columns)
{
    MPI_Datatype square_block;
    int stride = n;
    int count = mloc;
    int block_length = nloc;
    MPI_Type_vector(count, block_length, stride, MPI_INT, &square_block);
    MPI_Datatype square_block_resized;
    MPI_Type_create_resized(square_block, 0, sizeof(int), &square_block_resized);
    MPI_Type_commit(&square_block_resized);
    int *send_counts = (int*) calloc(world_size, sizeof(int));
    int *displs = (int*) calloc(world_size, sizeof(int));
    for (int i = 0; i < mesh_rows; i  ) {
        for (int j = 0; j < mesh_columns; j  ) {
            send_counts[i * mesh_columns   j] = 1;
            displs[i * mesh_columns   j] = i * n * block_length   j * block_length;
        }
    }
    
    MPI_Scatterv(Mat, send_counts, displs, square_block_resized, Mat_loc, mloc * nloc, MPI_INT, 0, MPI_COMM_WORLD);
}


bool is_divisible(int dividend, int divisor)
{
    return dividend % divisor == 0;
}

void build_mesh(MPI_Comm *mesh, MPI_Comm *mesh_r, MPI_Comm *mesh_c, int process_rank, int world_size,
    int mesh_rows, int mesh_columns, int *process_coordinates) 
{
    int mesh_dimension = 2;
    int *mesh_n_dimension;
    int mesh_reorder = 0;
    int *mesh_period;
    int *remain_dims = (int*) calloc(mesh_dimension, sizeof(int));
    mesh_n_dimension = (int*) calloc(mesh_dimension, sizeof(int));
    mesh_n_dimension[0] = mesh_rows;
    mesh_n_dimension[1] = mesh_columns;
    mesh_period = (int*) calloc(mesh_dimension, sizeof(int));
    mesh_period[0] = mesh_period[1] = 0;
    MPI_Cart_create(MPI_COMM_WORLD, mesh_dimension, mesh_n_dimension, mesh_period, mesh_reorder, mesh);
    MPI_Cart_coords(*mesh, process_rank, mesh_dimension, process_coordinates);
    remain_dims[0] = 0;  
    remain_dims[1] = 1;
    MPI_Cart_sub(*mesh, remain_dims, mesh_r);
    remain_dims[0] = 1;
    remain_dims[1] = 0;
    MPI_Cart_sub(*mesh, remain_dims, mesh_c);
}

int *fill_matrix(int *Mat, int m, int n)
{
    int k = 0;
    Mat = (int*) calloc(m * n, sizeof(int));
    for (int i = 0; i < m; i  )
        for (int j = 0; j < n; j  ) 
            Mat[i * n   j] =   k;
    return Mat;
}

As you can see this works fine but I wish I could re-write that commented part so I can exploit mesh_r communicator and distribute A_loc over every processors on mesh_r, instead of the hard-coded send with dest = 1 and dest = 2 over MPI_COMM_WORLD.

• Any help?

CodePudding user response：

Instead of sends and receives, you should use a Bcast as you did in an earlier version of your code. Your problem is that you're not thinking in a distributed manner, but you try to keep a global view. By that I mean that after you create the sub-communicator mesh_r, every process seems to be in that communicator, but and here it comes: there are multiple mesh_r communicators, and each process is part of exactly one. Each MPI process sees exactly the one mesh_r communicator that it is part of. Thus a single code line MPI_Bcast( ...buffer stuff...., mesh_r ) does multiple broadcasts, one in each grid row.