Question

我刚刚开始使用MPI编程，并决定使用OpenMPI创建一个简单的分布式qsort。要分发我要排序的数组的部分，我正在尝试使用MPI_Scatterv，但以下代码段错误对我：

#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;
}

我如何使用Scatterv？我是否在从工作程序代码中分配输入缓冲区时出错？

Answer 1

我更改了代码中的某些部分以使某些工作正常。

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个整数，但sizes[0]要小得多。其他进程还存在其他问题：MPI_Scatter将发送sizes[my_rank]个整数，但预计会sizes[my_rank-1] ...

这是一个将data_in从0分散到所有处理器的代码，包括0.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，应在代码末尾释放data_in和data_out。

这是你想做的吗？祝qsort好运！我认为你不是第一个使用MPI对整数进行排序的人。见parallel sort using mpi。你在0 processus上生成随机数然后分散它们的方法是正确的方法。我想你会被他的TD_Trier()通讯功能所吸引。即使您更改了tri_fusion(T, 0, size - 1);的{{1}} ...

再见，

弗朗西斯

MPI_Scatterv段错误

1 个答案: