Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

From: Chris Johnson (crjjrc_at_[hidden])
Date: 2007-10-02 14:38:05


Hi, I'm trying to run an MPI program of mine under OpenMPI 1.2 using
just one process (mpirun -np 1 ./a.out) and I'm getting some
unexpected results. The problem is that I'm getting unexpected
results from an MPI_Gatherv call when the offset for rank is nonzero.

I've worked up a small sample that can reproduce the problem on the
several machines I've tried. Here, each process creates a tmp array
of five ints. These tmp arrays are then gathered by rank into a
buffer, but offset by 10 places. (These 10 places are initialized
with -1.) When I run with multiple processes, I see the 10 -1s and
each process's tmp array in the buffer. But when I run with one
process, the buffer contains funny values. When I run with one
process under MPICH, the buffer contains the 10 -1s and the rank's
array, as expected. When the offset is 0, OpenMPI behaves just fine
with one process.

Here's the sample:

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

#define COUNT 5
#define OFFSET 10

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

   int i;
   int *nitems;
   int *offsets;
   int *buffer;
   int tmp[COUNT];
   int rank;
   int nprocs;

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

   for (i = 0; i < COUNT; i++) {
      tmp[i] = i + rank * 100;
   }

   if (rank == 0) {
      buffer = malloc(sizeof(int) * (nprocs * COUNT + OFFSET));
      nitems = malloc(sizeof(int) * nprocs);
      offsets = malloc(sizeof(int) * nprocs);
      nitems[0] = COUNT;
      offsets[0] = OFFSET;
      for (i = 1; i < nprocs; i++) {
         nitems[i] = COUNT;
         offsets[i] = offsets[i - 1] + nitems[i - 1];
      }

      for (i = 0; i < OFFSET; i++) {
         buffer[i] = -1;
      }
   }
   MPI_Gatherv(tmp, COUNT, MPI_INT, buffer, nitems, offsets, MPI_INT, 0,
               MPI_COMM_WORLD);

   if (rank == 0) {
      for (i = 0; i < nprocs * COUNT + OFFSET; i++) {
         printf("buffer[%d]: %d\n", i, buffer[i]);
      }
      free(buffer);
      free(nitems);
      free(offsets);
   }

   MPI_Finalize();

   return 0;

}
----------------------------------------------

For what it's worth, I've started using MPI_IN_PLACE instead of the
above method. This works around the problem for now, but I'd
appreciate any insight on how to fix this or confirmation of bug.
Thanks for your help!

- Chris