CODEWITHSUNDEEP
cparallel-processingmpiwritetofile
Amanda
Amanda

Reputation: 13

MPI_Gatherv is not collecting data correctly

I'm trying to gather different size arrays using MPI_Gatherv, but for some reason, it's only gathering the first object in the first processor. When I do the above for loop, I get the correct values from xPos and yPos, but when I gather the data to the xFinal and yFinal array and print out the values, I'm only getting the first x and y. So basically the first object has an (x,y) of (0,0) and I have 10 objects, and all the objects only print out (0,0) when the actual object it should be referencing has a different (x,y)

Just in case to say, the counts[rank] and displs are definitely right because I used them to scatterv previously.

Am I using gatherrv incorrectly? Or am I printing incorrectly?

for ( a = 0; a < size; a++) {
    if (rank == a) {
        for ( i = 0 ; i < counts[rank]; i++) {
            printf("from procs %d: %lE %lE\n", rank, xPos[i], yPos[i]);
        }
    }
}


MPI_Gatherv(&xPos, counts[rank], MPI_DOUBLE, &xFinal, counts, displs, MPI_DOUBLE,0, MPI_COMM_WORLD);
MPI_Gatherv(&yPos, counts[rank], MPI_DOUBLE, &yFinal, counts, displs, MPI_DOUBLE,0, MPI_COMM_WORLD);


MPI_Finalize();

FILE* f = fopen("universe.out", "wt");
for (i = 0; i < N; i++)
    fprintf(f, "%lE %lE\n", xFinal[i], yFinal[i]);
fclose(f);

Upvotes: 1

Views: 302

Answers (2)

Rob Latham
Rob Latham

Reputation: 5223

In my usual role of MPI-IO advocate, please consider MPI-IO for this kind of problem. You may be able to skip the gather entirely by having each process write to the file. Furthermore, if N is large, you do not want N file operations. Let MPI (either directly or through a library) make your life easier.

First off, is a text-based output format really what you (and your collaborators) want? if universe.out gets large enough that you want to read it in parallel, you are going to have a challenge decomposing the file across processors. Consider parallel HDF5 (phdf5) or paralllel-netcdf (pnetcdf) or any other of the higher-level libraries with a self-describing portable file format.

Here's an example of how you would write out all the x values then all the y values. 

#include <stdio.h>
#include <mpi.h>
#include <unistd.h> //getpid
#include <stdlib.h> //srandom, random

#define MAX_PARTS 10

int main(int argc, char **argv) {
    int rank, nprocs;
    MPI_Offset nparts; // more than 2 billion particle possible
    int i;
    double *x, *y;
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Offset start=0, x_end;
    MPI_Info info;
    MPI_File fh;
    MPI_Status status;
    srandom(getpid());

    /* demonstrate this works even when particles are not evenly
 * distributed over the processes */
    nparts=((double)random()/RAND_MAX)*MAX_PARTS;

    x = malloc(nparts*sizeof(*x));
    y = malloc(nparts*sizeof(*y));
    for (i = 0; i< nparts; i++) {
        /* just some bogus data to see what happens */
        x[i] = rank*100+i;
        y[i] = rank*200+i;
    }

    /* not using this now. might tune later if needed */
    MPI_Info_create(&info);

    MPI_File_open(MPI_COMM_WORLD, "universe.out",
    MPI_MODE_CREATE|MPI_MODE_WRONLY, info,  &fh);
    MPI_File_set_view(fh, 0, MPI_DOUBLE, MPI_DOUBLE, "native", info);

    MPI_Scan(&nparts, &start, 1, MPI_OFFSET, MPI_SUM, MPI_COMM_WORLD);

    /* MPI_Scan is a prefix reduction: remove our contribution */
    x_end = start; /* only the last rank will use this in the bcast below */
    start -= nparts;

    MPI_Bcast(&x_end, 1, MPI_OFFSET, nprocs-1, MPI_COMM_WORLD);
    MPI_File_write_at_all(fh, start, x, nparts, MPI_DOUBLE, &status);
    MPI_File_write_at_all(fh, start+x_end, y, nparts, MPI_DOUBLE, &status);

    MPI_Info_free(&info);
    MPI_File_close(&fh);

    MPI_Finalize();

}

Writing out the x array and y array as pairs of (x,y) values is possible, but a bit more complicated (You would have to make an MPI datatype).

Parallel-NetCDF has an "operation combining" optimization that does this for you. Parallel-HDF5 has a "multi-dataset i/o" optimization in the works for the next release. With those optimizations you could define a 3d array, with one dimension for x and y and the third for "particle identifier". Then you could post the operation for the x values, the operation for the y values, and let the library stich all that together into one call.

Upvotes: 1

Hristo Iliev
Hristo Iliev

Reputation: 74355

It seems that you are writing the file from all ranks simultaneously. You should put the file writing code within if (rank == 0) { ... } to only let rank 0 write:

if (rank == 0)
{
    FILE* f = fopen("universe.out", "wt");
    for (i = 0; i < N; i++)
        fprintf(f, "%lE %lE\n", xFinal[i], yFinal[i]);
    fclose(f);
}

Otherwise the content of the file could be anything.

Upvotes: 2

Related Questions