任意域的

Question

我有一个3 x 2的矩形域，如图（a）所示。使用MPI笛卡尔拓扑，我可以识别每个单元的4个邻居（-1表示NULL），如下所示

Figure: Rectangular and arbitrary domain

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

int main(int argc, char** argv) {
    int rank, size;
    int top = 0, right = 1, bottom = 2, left = 3;
    int neighbors[4], dimSize[2] = {3, 2};
    int usePeriods[2] = {0, 0}, newCoords[2];
    MPI_Comm cartComm = MPI_COMM_WORLD;

    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    if (size != 6) {
        if (rank == 0)
            printf("Please Launch with ONLY 6 processes  \n");
        MPI_Finalize();
        exit(0);
    }

    // Create a carthesian communicator
    MPI_Cart_create(MPI_COMM_WORLD, 2, dimSize, usePeriods, 1, &cartComm);

    // Obtain the 2D coordinates in the new communicator
    MPI_Cart_coords(cartComm, rank, 2, newCoords);

    // Obtain the direct neighbor ranks
    MPI_Cart_shift(cartComm, 0, 1, neighbors + left, neighbors + right);
    MPI_Cart_shift(cartComm, 1, 1, neighbors + top, neighbors + bottom);

    printf("Rank: %d \t Neighbors(top, right, bottom, left): %2d, %2d, %2d, %2d \n", rank, neighbors[0], neighbors[1], neighbors[2], neighbors[3]);

    MPI_Finalize();
    return 0;
}

结果：

Rank: 0      Neighbors(top, right, bottom, left): -1,  2,  1, -1
Rank: 1      Neighbors(top, right, bottom, left):  0,  3, -1, -1
Rank: 2      Neighbors(top, right, bottom, left): -1,  4,  3,  0
Rank: 3      Neighbors(top, right, bottom, left):  2,  5, -1,  1
Rank: 4      Neighbors(top, right, bottom, left): -1, -1,  5,  2
Rank: 5      Neighbors(top, right, bottom, left):  4, -1, -1,  3

我的问题：是否存在使用MPI识别非矩形域的这些邻居的方法，如图（b）所示。在这里，我们假设没有周期性邻居。现在，我只能想到为3x3实现类似的笛卡尔拓扑方法（图c）：

mpirun -np 9 ./run_cartesian

接下来，删除NULL单元格（灰色）并从0-5重新排序。将结果保存到文件，然后加载文件（使用root）并将相应的邻居分散到每个处理器。

mpirun -np 6 ./run_load_scatter

有关更好方法的建议吗？感谢。