Open MPI 的矩阵乘法错误

Question

我正在尝试使用 OpenMPI 和 C 计算 NxN 矩阵乘法。除MPI_Bcast() 外，一切都按预期运行。据我了解，MASTER 必须将matrix_2 广播给其余的 WORKER 进程。同时，当 WORKERS 到达MPI_Bcast() 时，他们应该在那里等待，直到选定的进程（在本例中为 MASTER）进行广播。

我得到的错误是分段错误和地址未映射，所以它肯定与矩阵的动态分配有关。我所做的是将matrix_1 的一部分发送到每个进程，然后每个进程与先前广播的matrix_2 进行部分乘法和加法运算。

我知道错误必须在MPI_Bcast() 上，因为当我评论它时，程序正确完成（但显然没有计算乘积）。一定有什么我不知道的。我留下了代码和收到的错误消息。提前致谢。

代码

#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
/* MACROS */
#define MASTER_TO_SLAVE_TAG 1
#define SLAVE_TO_MASTER_TAG 4 
#define MASTER 0 
#define WORKER 1

int *matrix_1;
int *matrix_2;

int *result;

double start_time;
double end_time;

int procID;
int numProc; 

int size, numRows, from, to;
int i,j,k;

MPI_Status status;
MPI_Request request;

void addressMatrixMemory(int);

int main(int argc, char *argv[]){
    size = atoi(argv[1]);
    MPI_Init (&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &procID); 
    MPI_Comm_size(MPI_COMM_WORLD, &numProc); 

    addressMatrixMemory(size);

    /* MASTER starts. */
    if(procID == MASTER){

        start_time = MPI_Wtime();
        for(i = 1; i < numProc; i++){
            numRows = size/(numProc - 1); 

            from = (i - 1) * numRows; 

            if(((i + 1) == numProc) && ((size % (numProc - 1))) != 0){
                to = size; 
            } else {
                to = from + numRows; 
            }
            MPI_Isend(&from, 1, MPI_INT, i, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &request);

            MPI_Isend(&to, 1, MPI_INT, i, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &request);

            MPI_Isend(matrix_1, (to - from) * size, MPI_INT, i, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &request);

        }

    }

    MPI_Bcast(&matrix_2, size * size, MPI_INT, MASTER, MPI_COMM_WORLD);

    /* WORKERS task */

    if(procID >= WORKER){
        int row, col;
        int *matrix = malloc(sizeof(matrix_1[0])*size*size);

        MPI_Recv(&from, 1, MPI_INT, MASTER, MASTER_TO_SLAVE_TAG, MPI_COMM_WORLD, &status);

        MPI_Recv(&to, 1, MPI_INT, MASTER, MASTER_TO_SLAVE_TAG + 1, MPI_COMM_WORLD, &status); 

        MPI_Recv(matrix, (to - from) * size, MPI_INT, MASTER, MASTER_TO_SLAVE_TAG + 2, MPI_COMM_WORLD, &status);
        for(row = from; row < to; row++){
            for(col = 0; col < size; col++){
                result[row * size + col] = 0;
                for(k = 0; k < size; k++);
                    result[row * size + col] += matrix[row * size + k] * matrix_2[k * size + col];
            }
        }
        MPI_Isend(&from, 1, MPI_INT, MASTER, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD, &request);
        MPI_Isend(&to, 1, MPI_INT, MASTER, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD, &request);
        MPI_Isend(&result[from], (to - from) * size, MPI_INT, MASTER, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD, &request);
    }

    /* MASTER gathers WORKERS job. */
    if(procID == MASTER){
        for(i = 1; i < numProc; i++){
            MPI_Recv(&from, 1, MPI_INT, i, SLAVE_TO_MASTER_TAG, MPI_COMM_WORLD, &status);
            MPI_Recv(&to, 1, MPI_INT, i, SLAVE_TO_MASTER_TAG + 1, MPI_COMM_WORLD, &status);
            MPI_Recv(&result[from], (to - from) * size, MPI_INT, i, SLAVE_TO_MASTER_TAG + 2, MPI_COMM_WORLD, &status);
        }

        end_time = MPI_Wtime();
        printf("\nRunning Time = %f\n\n", end_time - start_time);
    }

    MPI_Finalize();

    free(matrix_1);
    free(matrix_2);
    free(result);

    return EXIT_SUCCESS;
}

void addressMatrixMemory(int n){
    matrix_1 = malloc(sizeof(matrix_1[0])*n*n);
    matrix_2 = malloc(sizeof(matrix_2[0])*n*n);

    result = malloc(sizeof(result[0])*n*n);

    /* Matrix init with values between 1 y 100. */
    srand(time(NULL));
    int r = rand() % 100 + 1;
    int i;
    for(i = 0; i < n*n; i++){
        matrix_1[i] = r;
        r = rand() % 100 + 1;
        matrix_2[i] = r;
        r = rand() % 100 + 1;
    }
}

错误信息

[tuliansPC:28270] *** Process received signal ***
[tuliansPC:28270] Signal: Segmentation fault (11)
[tuliansPC:28270] Signal code: Address not mapped (1)
[tuliansPC:28270] Failing at address: 0x603680
    [tuliansPC:28270] [ 0] /lib/x86_64-linux-gnu/libpthread.so.0(+0x10340) [0x7f0a98ce0340]
[tuliansPC:28270] [ 1] /lib/x86_64-linux-gnu/libc.so.6(+0x97ffe) [0x7f0a9899fffe]
[tuliansPC:28270] [ 2] /usr/lib/libmpi.so.1(opal_convertor_pack+0x129) [0x7f0a98fef779]
[tuliansPC:28270] [ 3] /usr/lib/openmpi/lib/openmpi/mca_btl_sm.so(mca_btl_sm_prepare_src+0x1fd) [0x7f0a923c385d]
[tuliansPC:28270] [ 4] /usr/lib/openmpi/lib/openmpi/mca_pml_ob1.so(mca_pml_ob1_send_request_start_rndv+0x1dc) [0x7f0a93245c9c]
[tuliansPC:28270] [ 5] /usr/lib/openmpi/lib/openmpi/mca_pml_ob1.so(mca_pml_ob1_isend+0x8ec) [0x7f0a9323856c]
[tuliansPC:28270] [ 6] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_generic+0x3fc) [0x7f0a914f49fc]
[tuliansPC:28270] [ 7] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_pipeline+0xbc) [0x7f0a914f4d5c]
[tuliansPC:28270] [ 8] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_dec_fixed+0x134) [0x7f0a914ec7a4]
[tuliansPC:28270] [ 9] /usr/lib/openmpi/lib/openmpi/mca_coll_sync.so(mca_coll_sync_bcast+0x64) [0x7f0a917096a4]
[tuliansPC:28270] [10] /usr/lib/libmpi.so.1(MPI_Bcast+0x13d) [0x7f0a98f5678d]
[tuliansPC:28270] [11] ej5Exec() [0x400e8c]
[tuliansPC:28270] [12] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf5) [0x7f0a98929ec5]
[tuliansPC:28270] [13] ej5Exec() [0x400ac9]
[tuliansPC:28270] *** End of error message ***
--------------------------------------------------------------------------
mpirun noticed that process rank 0 with PID 28270 on node tuliansPC exited on signal 11 (Segmentation fault).
--------------------------------------------------------------------------

Answer 1

让我们从跳出来的第一个问题开始。您错误地使用了非阻塞通信。 MPI_Isend 是一个非阻塞发送函数，这意味着当您调用 MPI_Isend 时，您真正要做的只是告诉 MPI 您想在未来某个时间发送的消息。它可能会立即发送，也可能不会。为了保证数据真的被发送，你需要用MPI_Wait之类的东西来完成调用。通常当人们使用非阻塞调用（MPI_Isend）时，他们不会将它们与阻塞调用（@987654325 @)。如果您使用所有非阻塞调用，那么您可以使用一个函数 MPI_Waitall 完成所有这些调用。

尝试先解决这些问题，看看是否能解决您的问题。仅仅因为您评论了集体，并不意味着其他问题不存在。由于这种奇怪的行为，MPI 程序是出了名的难以调试。