Minor Issue With Alltoall

David_Race — Mon, 19 Dec 2011 18:34:22 GMT

The work-around for this is rather obvious, but the behavior between MPI_Alltoall and MPI_Allgather is different as seen in the attached code. According to the standard for both -

The "in place" option for intracommunicators is specied by passing the value

MPI_IN_PLACE to the argument sendbuf at all processes. sendcount and sendtype are ignored.

The \\in place" option for intracommunicators is specied by passing the value MPI_IN_PLACE to the argument sendbuf at all processes. sendcount and sendtype are ignored.

As seen in the code, this works correctly for MPI_Allgather. For MPI_Alltoall, the following error occurs

Fatal error in MPI_Alltoall: Invalid datatype, error stack:

MPI_Alltoall(851): MPI_Alltoall(sbuf=MPI_IN_PLACE, scount=0, MPI_DATATYPE_NULL, rbuf=0x7fffbfbaa410, rcount=1, MPI_INT, MPI_COMM_WOR

LD) failed

MPI_Alltoall(740): Null Datatype pointer

The MPI_DATATYPE_NULL should be ignored along with the scount.

// test program for the MPI_Type_size bug

// David Race

// 17 December 2011

#include

// mpi

#include "mpi.h"

#define MAX_SIZE 32

int main ( int argc, char *argv[] )

{

// definitions

int mpierror, isize, myRank;

int typeSize;

int valA[MAX_SIZE];

int i;

int commRoot;

// start processing

printf("Start of program\\n");

printf("SIZE OF VALA %ld\\n",sizeof(valA));

mpierror = MPI_Init ( &argc, &argv );

mpierror = MPI_Comm_rank ( MPI_COMM_WORLD, &myRank );

mpierror = MPI_Comm_size ( MPI_COMM_WORLD, &isize );

MPI_Barrier(MPI_COMM_WORLD);

// test the mpi_type_size using MPI_Allgather

if (myRank == 0) {

printf("=====================================================\\n");

printf(" Allgather \\n");

printf("=====================================================\\n");

fflush(stdout);

}

for(i=0;i = 0;

valA[myRank] = myRank;

mpierror = MPI_Allgather(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, valA, 1, MPI_INT, MPI_COMM_WORLD);

if (myRank == 0) {

for(i=0;i);

fflush(stdout);

}

MPI_Barrier(MPI_COMM_WORLD);

// test the mpi_type_size using MPI_Alltoall

if (myRank == 0) {

printf("=====================================================\\n");

printf(" Alltoall \\n");

printf("=====================================================\\n");

fflush(stdout);

}

for(i=0;i = i;

commRoot = 0;

mpierror = MPI_Alltoall(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, valA, 1, MPI_INT, MPI_COMM_WORLD);

if (myRank == 0) {

for(i=0;i);

fflush(stdout);

}

MPI_Barrier(MPI_COMM_WORLD);

if (myRank == 1) {

for(i=0;i);

fflush(stdout);

}

MPI_Barrier(MPI_COMM_WORLD);

mpierror = MPI_Finalize();

fflush(stdout);

printf("End of Program\\n");

}

Minor Issue With Alltoall

Dmitry_S_Intel — Tue, 20 Dec 2011 10:41:05 GMT

Hi David,

Thank you for the post. We have submitted the issue to our tracker.

--
Dmitry

topic Minor Issue With Alltoall in Intel® MPI Library

Minor Issue With Alltoall

Minor Issue With Alltoall