2.0.0rc2: Regression in MPI_Ireduce_scatter_block()

[dalcinl]

The following trivial test checks that MPI_Reduce_scatter_block() and MPI_Ireduce_scatter_block() produce the expected output and do not overwrite past the receive buffer. To keep it simple, it can be run in at most 16 processes.

Using Open MPI 1.10.2, I get the expected output in all cases, so this is definitely a regression in 2.0.0rc2.

#include <stdlib.h>
#include <stdio.h>
#include <mpi.h>
int main(int argc, char *argv[])
{
  int i,size,rank;
  int sendbuf[] = {1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1};
  int recvbuf[] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
  int rcount    = 1;

  MPI_Init(&argc, &argv);
  MPI_Comm_size(MPI_COMM_WORLD, &size);
  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
  if (size > 16) MPI_Abort(MPI_COMM_WORLD,1);
#ifndef NBCOLL
#define NBCOLL 1
#endif
#if NBCOLL
  MPI_Request request;
  MPI_Ireduce_scatter_block(sendbuf, recvbuf, rcount, MPI_INT,
                            MPI_SUM, MPI_COMM_WORLD, &request);
  MPI_Wait(&request,MPI_STATUS_IGNORE);
#else
  MPI_Reduce_scatter_block(sendbuf, recvbuf, rcount, MPI_INT,
                           MPI_SUM, MPI_COMM_WORLD);
#endif
  printf("[%d] rbuf[%d]=%2d  expected:%2d\n", rank, 0, recvbuf[0], size);
  for (i=1; i<size+1; i++) {
    printf("[%d] rbuf[%d]=%2d  expected:%2d\n", rank, i, recvbuf[i], 0);
  }
  MPI_Finalize();
  return 0;
}

Using MPI_Reduce_scatter_block(), I get the expected output for any number of processes

$ mpicc -DNBCOLL=0 ireduce_scatter_block.c

$ mpiexec -n 1 ./a.out 
[0] rbuf[0]= 1  expected: 1
[0] rbuf[1]= 0  expected: 0

$ mpiexec -n 3 ./a.out 
[0] rbuf[0]= 3  expected: 3
[0] rbuf[1]= 0  expected: 0
[0] rbuf[2]= 0  expected: 0
[0] rbuf[3]= 0  expected: 0
[1] rbuf[0]= 3  expected: 3
[1] rbuf[1]= 0  expected: 0
[1] rbuf[2]= 0  expected: 0
[1] rbuf[3]= 0  expected: 0
[2] rbuf[0]= 3  expected: 3
[2] rbuf[1]= 0  expected: 0
[2] rbuf[2]= 0  expected: 0
[2] rbuf[3]= 0  expected: 0

Using MPI_Ireduce_scatter_block(), the output for one process P=1 is wrong, but for P>1 we are good:

$ mpicc -DNBCOLL=1 ireduce_scatter_block.c

$ mpiexec -n 1 ./a.out 
[0] rbuf[0]= 0  expected: 1
[0] rbuf[1]= 0  expected: 0

$ mpiexec -n 3 ./a.out 
[0] rbuf[0]= 3  expected: 3
[0] rbuf[1]= 0  expected: 0
[0] rbuf[2]= 0  expected: 0
[0] rbuf[3]= 0  expected: 0
[1] rbuf[0]= 3  expected: 3
[1] rbuf[1]= 0  expected: 0
[1] rbuf[2]= 0  expected: 0
[1] rbuf[3]= 0  expected: 0
[2] rbuf[0]= 3  expected: 3
[2] rbuf[1]= 0  expected: 0
[2] rbuf[2]= 0  expected: 0
[2] rbuf[3]= 0  expected: 0

[ggouaillardet]

the fix got lost when libnbc was rewritten ...
I will fix that tomorrow

[ggouaillardet]

Fixed in open-mpi/ompi-release#1143