3

我是 MPI 的新手，我遇到了这个问题。我想读取一个超过 20000 行的文件的内容，然后将这些行平均分配给所有进程以进行进一步处理。文件中每一行的内容是这样的（两列数字），

45.87   13.22
45.71   13.27
45.78   13.21
45.67   13.1
45.7    13.24
45.81   13.28
45.85   13.32

我需要在运行时将行平均划分为任意数量的进程（进程数可以是例如；2、3、4、5、....、128）

我知道如何将文件分成一个块，但我需要保留每一行中的值，所以我需要逐行读取。

这是我用来完成这项工作的 MPI 代码和串行代码，但我遇到了分段错误。

/* Open the file */
MPI_File_open (MPI_COMM_WORLD, "small.txt", MPI_MODE_RDONLY, MPI_INFO_NULL, &myfile);
/* Get the size of the file */
MPI_File_get_size(myfile, &filesize);
/* Calculate how many elements that is */
filesize = filesize/sizeof(char);

/* Calculate how many elements each processor gets */
bufsize = filesize/np;
/* Allocate the buffer to read to, one extra for terminating null char */
buf = (char *) malloc((bufsize+1)*sizeof(char));


/* Set the file view */
MPI_File_set_view(myfile, myid*bufsize*sizeof(char), MPI_CHAR, MPI_CHAR,"native",MPI_INFO_NULL);


Nooflines_Real = count_lines(myfile);
printf("%s contains %d lines\n", argv[1], Nooflines_Real);


int count_lines (FILE *infile) {
  char readline[80];
  int lines=0;
  while( fgets(readline,80,infile) != NULL ) lines++;
  rewind(infile);
  return(lines);
}
4

1 回答 1

4

您的参数myfile是 type 的变量MPI_File，而不是FILE *，因此您不能将其用于fgets(),rewind()等。这可能是您的段错误的来源。

我的建议是采用此答案中的方法并读取每个文件的重叠块（以考虑到您不知道一行有多长的事实），每个任务都读取它们的块和进程their行。如果您真的关心每个文件具有完全相同的行数（尽可能），您可以让它们相互交换数据以具有完全相同的行数。

更新：如果你真的想这样做（请注意，如果你的输入都是数字，这在二进制格式中会容易得多），一些读取文本文件的代码，在另一个数字中的分区，然后处理每个行（例如通过对列求和）作为我上面链接的答案的直接扩展：

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

void readlines(MPI_File *in, const int rank, const int size, const int overlap,
               char ***lines, int *nlines) {
    MPI_Offset filesize;
    MPI_Offset localsize;
    MPI_Offset start;
    MPI_Offset end;
    char *chunk;

    /* figure out who reads what */

    MPI_File_get_size(*in, &filesize);
    localsize = filesize/size;
    start = rank * localsize;
    end   = start + localsize - 1;

    /* add overlap to the end of everyone's chunk... */
    end += overlap;

    /* except the last processor, of course */
    if (rank == size-1) end = filesize;

    localsize =  end - start + 1;

    /* allocate memory */
    chunk = malloc( (localsize + 1)*sizeof(char));

    /* everyone reads in their part */
    MPI_File_read_at_all(*in, start, chunk, localsize, MPI_CHAR, MPI_STATUS_IGNORE);
    chunk[localsize] = '\0';

    /*
     * everyone calculate what their start and end *really* are by going 
     * from the first newline after start to the first newline after the
     * overlap region starts (eg, after end - overlap + 1)
     */

    int locstart=0, locend=localsize;
    if (rank != 0) {
        while(chunk[locstart] != '\n') locstart++;
        locstart++;
    }
    if (rank != size-1) {
        locend-=overlap;
        while(chunk[locend] != '\n') locend++;
    }
    localsize = locend-locstart+1;

    /* Now let's copy our actual data over into a new array, with no overlaps */
    char *data = (char *)malloc((localsize+1)*sizeof(char));
    memcpy(data, &(chunk[locstart]), localsize);
    data[localsize] = '\0';
    free(chunk);

    /* Now we'll count the number of lines */
    *nlines = 0;
    for (int i=0; i<localsize; i++)
        if (data[i] == '\n') (*nlines)++;

    /* Now the array lines will point into the data array at the start of each line */
    /* assuming nlines > 1 */
    *lines = (char **)malloc((*nlines)*sizeof(char *));
    (*lines)[0] = strtok(data,"\n");
    for (int i=1; i<(*nlines); i++)
        (*lines)[i] = strtok(NULL, "\n");

    return;
}

void processlines(char **lines, const int nlines, const int rank) {
    for (int i=0; i<nlines; i++) {
        float a, b;
        sscanf(lines[i],"%f %f", &a, &b);
        printf("%d: <%s>: %f + %f = %f\n", rank, lines[i], a, b, a+b);
    }
}

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

    MPI_File in;
    int rank, size;
    int ierr;

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

    if (argc != 2) {
        if (rank == 0) fprintf(stderr, "Usage: %s infilename\n", argv[0]);
        MPI_Finalize();
        exit(1);
    }

    ierr = MPI_File_open(MPI_COMM_WORLD, argv[1], MPI_MODE_RDONLY, MPI_INFO_NULL, &in);
    if (ierr) {
        if (rank == 0) fprintf(stderr, "%s: Couldn't open file %s\n", argv[0], argv[1]);
        MPI_Finalize();
        exit(2);
    }

    const int overlap=100;
    char **lines;
    int nlines;
    readlines(&in, rank, size, overlap, &lines, &nlines);

    printf("Rank %d has %d lines\n", rank, nlines);

    processlines(lines, nlines, rank);

    free(lines[0]);
    free(lines);

    MPI_File_close(&in);

    MPI_Finalize();
    return 0;
}

并在您提供的数据集上运行它：

$ mpirun -np 2 ./textio foo2.in 
Rank 0 has 4 lines
0: <45.87   13.22>: 45.869999 + 13.220000 = 59.090000
0: <45.71   13.27>: 45.709999 + 13.270000 = 58.980000
0: <45.78   13.21>: 45.779999 + 13.210000 = 58.989998
0: <45.67   13.1>: 45.669998 + 13.100000 = 58.769997
Rank 1 has 3 lines
1: <45.7    13.24>: 45.700001 + 13.240000 = 58.940002
1: <45.81   13.28>: 45.810001 + 13.280000 = 59.090000
1: <45.85   13.32>: 45.849998 + 13.320000 = 59.169998
于 2012-11-11T05:23:57.270 回答