在简单MPI代码中读写测试文件时出错

Question

我有一个MPI代码，其中进程读取二进制文件并将其重新写回。数据的分布方式是，进程0读取(然后写入)文件的前半部分，而进程1读取(然后写入)文件的后半部分。现在的问题是输入和输出文件不匹配(diff表明它们不同)。如果只有一个进程，则一切正常。有人能指出哪里出了问题吗？

使用OpenMPI，编译为: mpicc -Wall test_mpi.c -o test_mpi

运行方式: mpirun 2./ -np _mpi

提前谢谢。

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

int main(int argc, char** argv) {

    int rank, np, i;  //np = no. of processes
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &np);

    int filesize = 48*1048576;  //input filesize 48MB

    double *data = (double*) malloc (filesize/np);
    FILE* fpa;
    fpa = fopen ( "512_featurevec.out", "rb");
    fseek(fpa, filesize/np*rank, SEEK_SET);
    printf("read: %d\n", (int)fread(&data[0], sizeof(double), filesize/(np*sizeof(double)), fpa));
    fclose(fpa);

    char* outfile = "outfile.txt";
    for(i=0; i<np; i++) {
        if(rank == i) {
            fpa = fopen ( outfile, "ab");
            fseek(fpa, filesize/np*rank, SEEK_SET);
            fwrite ( &data[0], sizeof(double), filesize/(np*sizeof(double)), fpa);
            fclose ( fpa );
        }   
    }   

    free(data);
    MPI_Finalize();
    exit(0);
}

Answer 1

回答我自己的问题，但这是有效的：

char* outfile = "outfile.txt";
    for(i=0; i<np; i++) {
        if(rank == i) {
            fpa = fopen ( outfile, "ab");
            fseek(fpa, filesize/np*rank, SEEK_SET);
            fwrite ( &data[0], sizeof(double), filesize/(np*sizeof(double)), fpa);
            fclose ( fpa );
        } 
        MPI_Barrier(comm);  
    } 

出于某种原因，如果proc 1在proc 0之前到达，并将一个空文件查找到某个值，它会将文件指针定位到正确的偏移量(由ftell验证)，但仅从偏移量0开始写入。(我肯定做错了什么，但不管怎样)。

Answer 2

看起来这个问题是由于每个孩子都在打开文件进行写入，这导致了争用。

尝试让文件名取决于等级(例如，写入out file.txt.(rank)并查看是否所有输出都匹配)。

Answer 3

如果您已经在使用MPI，并且不厌其烦地使用seek对文件进行分区，而不是使用POSIX，我建议您使用MPI-IO (作为MPI2的标准部分，大约1996年左右)：

• http://www.mhpcc.edu/training/workshop2/mpi_io/MAIN.html
• http://beige.ucs.indiana.edu/B673/node179.html

在我们的中心，我们有这个的第一部分，我认为这是非常好的：

• https://support.scinet.utoronto.ca/wiki/images/a/aa/ParIO-HPCS2011.pdf

上面代码的MPI-IOed版本如下：

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

int main(int argc, char** argv) {

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

    const int filesize = 48*1048576;  //input filesize 48MB
    const int ndoubles = filesize/(sizeof(double)*np);
    double *localdata = malloc(ndoubles*sizeof(double));

    /* create a type which describes our view of the file --
     * in particular, just our subarray of the global array
     */
    int globalsizes[1] = {filesize};
    int localsizes[1] = {ndoubles};
    int starts[1] = {ndoubles*rank};

    MPI_Datatype fileview;
    MPI_Type_create_subarray(1, globalsizes, localsizes, starts, MPI_ORDER_C, MPI_DOUBLE, &fileview);
    MPI_Type_commit(&fileview);

    /* read in only our data */
    MPI_File fpa;
    MPI_Status status;
    MPI_File_open(MPI_COMM_WORLD, "512_featurevec.out", MPI_MODE_RDONLY, MPI_INFO_NULL, &fpa);

    /* note could use MPI_File_seek instead of file set view */
    MPI_File_set_view(fpa, (MPI_Offset)0, MPI_DOUBLE, fileview, "native", MPI_INFO_NULL);
    MPI_File_read_all(fpa, localdata, ndoubles, MPI_DOUBLE, &status);
    MPI_File_close(&fpa);

    /* write out data - it will have same layout, we're just writing instead of erading*/

    MPI_File_open(MPI_COMM_WORLD, "output.dat", MPI_MODE_WRONLY|MPI_MODE_CREATE, MPI_INFO_NULL, &fpa);

    /* note could use MPI_File_seek instead of file set view */
    MPI_File_set_view(fpa, (MPI_Offset)0, MPI_DOUBLE, fileview, "native", MPI_INFO_NULL);
    MPI_File_write_all(fpa, localdata, ndoubles, MPI_DOUBLE, &status);
    MPI_File_close(&fpa);

    free(localdata);
    MPI_Type_free(&fileview);
    MPI_Finalize();

    return 0;
}