当使用与处理器关联的不同工作负载时，MPI 会在发送后阻止执行

问：

我在使用 MPI 代码时遇到了一些问题（由我编写，用于测试另一个程序，其中不同的工作负载与不同的处理器相关联）。问题是，当我使用与 1 或 arraySize（在本例中为 4）不同数量的处理器时，程序在MPI_Send期间被阻塞，特别是当我运行程序时，程序在调用期间被阻塞。我现在没有使用任何调试器，我只是想了解为什么它适用于 1 个和 4 个处理器，但它不适用于 2 个处理器（每个处理器在阵列中有 2 个位置），代码如下：mpirun -np 2 MPItest

#include <mpi.h>
#include <iostream>

int main(int argc, char** argv) {
    int rank, size;
    const int arraySize = 4;
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    // every processor have a different workload (1 or more spots on the array to send to the other processors)
    // every processor sends to every other processor its designated spots


    int* sendbuf = new int[arraySize];
    int* recvbuf = new int[arraySize];

    int istart = arraySize/size * rank;
    int istop = (rank == size) ? arraySize : istart + arraySize/size;

    for (int i = istart; i < istop; i++) {
        sendbuf[i] = i;
    }

    std::cout << "Rank " << rank << " sendbuf :" << std::endl;
    //print the sendbuf before receiving its other values
    for (int i = 0; i < arraySize; i++) {
        std::cout << sendbuf[i] << ", ";
    }
    std::cout << std::endl;

    // sending designated spots of sendbuf to other processors
    for(int i = istart; i < istop; i++){
        for(int j = 0; j < size; j++){
            MPI_Send(&sendbuf[i], 1, MPI_INT, j, i, MPI_COMM_WORLD);
        }
    }

    // receiving the full array
    for(int i = 0; i < arraySize ; i++){
        int recvRank = i/(arraySize/size);
        MPI_Recv(&recvbuf[i], 1, MPI_INT, recvRank, i, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
    }


    // print the recvbuf after receiving its other values
    std::cout << "Rank " << rank << " recvbuf :" << std::endl;
    for (int i = 0; i < arraySize; i++) {
        std::cout << recvbuf[i] << ", ";
    }
    std::cout << std::endl;

    delete[] sendbuf;
    delete[] recvbuf;

    MPI_Finalize();
    return 0;
}

我正在使用标签来区分数组中的不同点（也许这就是问题所在？

我尝试使用不同数量的处理器，使用 1 个处理器程序可以工作，使用 4 个处理器也可以使用程序，使用 3 个处理器它会崩溃，使用 2 个处理器程序会被阻止。我也尝试使用MPI_Isend但它也不起作用（标志为 0），带有 MPI_Isend 的修改代码如下：

#include <mpi.h>
#include <iostream>

int main(int argc, char** argv) {
    int rank, size;
    const int arraySize = 4;
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    // every processor have a different workload (1 or more spots on the array to send to the other processors)
    // every processor sends to every other processor its designated spots


    int* sendbuf = new int[arraySize];
    int* recvbuf = new int[arraySize];

    int istart = arraySize/size * rank;
    int istop = (rank == size) ? arraySize : istart + arraySize/size;

    for (int i = istart; i < istop; i++) {
        sendbuf[i] = i;
    }

    std::cout << "Rank " << rank << " sendbuf :" << std::endl;
    //print the sendbuf before receiving its other values
    for (int i = 0; i < arraySize; i++) {
        std::cout << sendbuf[i] << ", ";
    }
    std::cout << std::endl;

    // sending designated spots of sendbuf to other processors
    for(int i = istart; i < istop; i++){
        for(int j = 0; j < size; j++){
            MPI_Request request;
            //MPI_Send(&sendbuf[i], 1, MPI_INT, j, i, MPI_COMM_WORLD);
            MPI_Isend(&sendbuf[i], 1, MPI_INT, j, i, MPI_COMM_WORLD, &request);
            // control if the send is completed
            int flag = 0;
            MPI_Test(&request, &flag, MPI_STATUS_IGNORE);
            const int numberOfRetries = 10;
            if(flag == 0){ // operation not completed
                std::cerr << "Error in sending, waiting" << std::endl;
                for(int k = 0; k < numberOfRetries; k++){
                    MPI_Test(&request, &flag, MPI_STATUS_IGNORE);
                    if(flag == 1){
                        break;
                    }
                }
                if(flag == 0){
                    std::cerr << "Error in sending, aborting" << std::endl;
                    MPI_Abort(MPI_COMM_WORLD, 1);
                }
                
            }
        }
    }

    // receiving the full array
    for(int i = 0; i < arraySize ; i++){
        int recvRank = i/(arraySize/size);
        MPI_Recv(&recvbuf[i], 1, MPI_INT, recvRank, i, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
    }


    // print the recvbuf after receiving its other values
    std::cout << "Rank " << rank << " recvbuf :" << std::endl;
    for (int i = 0; i < arraySize; i++) {
        std::cout << recvbuf[i] << ", ";
    }
    std::cout << std::endl;

  
    //MPI_Alltoall(sendbuf, 1, MPI_INT, recvbuf, 1, MPI_INT, MPI_COMM_WORLD);

    delete[] sendbuf;
    delete[] recvbuf;

    MPI_Finalize();
    return 0;
}

使用此代码，-np 4 也不起作用