c - MPI_Scatterv 段错误

Question

我刚刚开始使用 MPI 编程，并决定使用 OpenMPI 制作一个简单的分布式 qsort。为了分配我想要排序的数组部分，我正在尝试使用 MPI_Scatterv，但是以下代码对我造成段错误:

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

#define ARRAY_SIZE 26
#define BUFFER_SIZE 2048

int main(int argc, char** argv) {
    int my_rank, nr_procs;

    int* data_in, *data_out;
    int* sizes;
    int* offsets;

    srand(time(0));

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nr_procs);
    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);

    // everybody generates the control tables
    int nr_workers = nr_procs-1;
    sizes = malloc(sizeof(int)*nr_workers);
    offsets = malloc(sizeof(int)*nr_workers);

    int nr_elems = ARRAY_SIZE/nr_workers;

    // basic distribution
    for (int i = 0; i < nr_workers; ++i) {
        sizes[i] = nr_elems;
    }
    // distribute the remainder
    int left = ARRAY_SIZE%nr_workers;
    int curr_worker = 0;
    while (left) {
        ++sizes[curr_worker];
        curr_worker = (++curr_worker)%nr_workers;
        --left;
    }   
    // offsets
    int curr_offset = 0;
    for (int i = 0; i < nr_workers; ++i) {
        offsets[i] = curr_offset;
        curr_offset += sizes[i];
    }


    if (my_rank == 0) {
        // root
        data_in = malloc(sizeof(int)*ARRAY_SIZE);
        data_out = malloc(sizeof(int)*ARRAY_SIZE);
        for (int i = 0; i < ARRAY_SIZE; ++i) {
            data_in[i] = rand();
        }

        for (int i = 0; i < nr_workers; ++i) {
            printf("%d at %d\n", sizes[i], offsets[i]);
        }

        MPI_Scatterv (data_in, sizes, offsets, MPI_INT, data_out, ARRAY_SIZE, MPI_INT, 0, MPI_COMM_WORLD);


    } else {
        // worker
        printf("%d has %d elements!\n",my_rank, sizes[my_rank-1]);
        // alloc the input buffer
        data_in = malloc(sizeof(int)*sizes[my_rank-1]);

        MPI_Scatterv(NULL, NULL, NULL, MPI_INT, data_in, sizes[my_rank-1], MPI_INT, 0, MPI_COMM_WORLD);

        printf("%d got:\n", my_rank);
        for (int i = 0; i < sizes[my_rank-1]; ++i) {
            printf("%d ", data_in[i]);
        }
        printf("\n");
    }

    MPI_Finalize();
    return 0;
}

我将如何使用 Scatterv？从工作代码内部分配我的输入缓冲区是不是我做错了什么？

Answer 1

我更改了您代码中的某些部分以使其正常工作。

MPI_Scatter() 会将数据发送到每个处理器，包括他自己。根据您的程序，处理器 0 需要 ARRAY_SIZE 整数，但 sizes[0] 小得多。其他进程还有其他问题:MPI_Scatter 将发送 sizes[my_rank] 整数，但 sizes[my_rank-1] 将被期望...

这是将 data_in 从 0 分散到所有处理器(包括 0)的代码。因此我将 1 添加到 nr_workers :

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

#define ARRAY_SIZE 26
#define BUFFER_SIZE 2048

int main(int argc, char** argv) {
    int my_rank, nr_procs;

    int* data_in, *data_out;
    int* sizes;
    int* offsets;

    srand(time(0));

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nr_procs);
    MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);

    // everybody generates the control tables
    int nr_workers = nr_procs;
    sizes = malloc(sizeof(int)*nr_workers);
    offsets = malloc(sizeof(int)*nr_workers);

    int nr_elems = ARRAY_SIZE/nr_workers;

    // basic distribution
    for (int i = 0; i < nr_workers; ++i) {
        sizes[i] = nr_elems;
    }
    // distribute the remainder
    int left = ARRAY_SIZE%nr_workers;
    int curr_worker = 0;
    while (left) {
        ++sizes[curr_worker];
        curr_worker = (++curr_worker)%nr_workers;
        --left;
    }   
    // offsets
    int curr_offset = 0;
    for (int i = 0; i < nr_workers; ++i) {
        offsets[i] = curr_offset;
        curr_offset += sizes[i];
    }

    if (my_rank == 0) {
        // root
        data_in = malloc(sizeof(int)*ARRAY_SIZE);

        for (int i = 0; i < ARRAY_SIZE; ++i) {
            data_in[i] = rand();
            printf("%d %d \n",i,data_in[i]);
        }

        for (int i = 0; i < nr_workers; ++i) {
            printf("%d at %d\n", sizes[i], offsets[i]);
        }


    } else {
        printf("%d has %d elements!\n",my_rank, sizes[my_rank]);
    }
    data_out = malloc(sizeof(int)*sizes[my_rank]);

    MPI_Scatterv (data_in, sizes, offsets, MPI_INT, data_out, sizes[my_rank], MPI_INT, 0, MPI_COMM_WORLD);

    printf("%d got:\n", my_rank);
    for (int i = 0; i < sizes[my_rank]; ++i) {
        printf("%d ", data_out[i]);
    }
    printf("\n");
    free(data_out);
    if(my_rank==0){
        free(data_in);
    }
    MPI_Finalize();
    return 0;
}

关于内存管理，data_in 和 data_out 应该在代码末尾释放。

这是你想做的吗？祝 qsort 好运!我认为您不是第一个使用 MPI 对整数进行排序的人。参见 parallel sort using mpi .您在 0 进程上生成随机数然后分散它们的方法是正确的方法。我想你会对他的 TD_Trier() 通信功能感兴趣。即使您将 tri_fusion(T, 0, size - 1); 更改为 qsort(...)...

再见，

弗朗西斯