[mpich-discuss] MPI_Bsend

[Dave Goodell]

You need to add MPI_BSEND_OVERHEAD to your buffer size.  This is an upper bound on the amount of space that the MPI implementation will internally use from your buffer.  So if you want to guarantee that X bytes will be successfully buffered by the library then you need to attach a buffer of size X+MPI_BSEND_OVERHEAD.  Officially you are supposed to use MPI_Pack_size too, although I'm not sure if that's required in practical implementations.

See MPI-2.2, pages 47-48.

-Dave

On Nov 10, 2011, at 1:30 PM CST, Wei-keng Liao wrote:

> My program using MPI_Bsend failed due to insufficient space,
> even though it did allocate exact size required, but not one byte more.
> I wonder if this is an mpich bug, or MPI requires more space than needed.
> If it is later, how much more is needed?
> 
> Wei-keng
> 
> error message:
> $ mpiexec -machinefile=machinefile -l -n 4 a.out
> [2] Fatal error in MPI_Bsend: Invalid buffer pointer, error stack:
> [2] MPI_Bsend(182).......: MPI_Bsend(buf=0x7fff81048a50, count=1024, MPI_BYTE, dest=0, tag=2, MPI_COMM_WORLD) failed
> [2] MPIR_Bsend_isend(318): Insufficient space in Bsend buffer; requested 1024; total buffer size is 4096
> 
> ---- bsend.c -----------------------------
> #include <stdio.h>
> #include <stdlib.h>
> #include <mpi.h>
> 
> int main(int argc, char **argv)
> {
>    int  i, rank, nprocs, src, bufsize;
>    char buf[1024];
>    void *bsend_buf = NULL;
>    MPI_Status status;
> 
>    MPI_Init(&argc,&argv);
>    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
>    MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
> 
>    bufsize = 1024*4;
>    // bufsize += 512; // if uncommented, the program runs successfully
>    MPI_Buffer_attach(malloc(bufsize), bufsize);
> 
>    for (i=0; i<4; i++) {
>        if (rank > 0)
>            MPI_Bsend(buf, 1024, MPI_BYTE, 0, rank, MPI_COMM_WORLD);
>        else {
>            for (src=1; src<nprocs; src++)
>                MPI_Recv(buf, 1024, MPI_BYTE, src, src, MPI_COMM_WORLD, &status);
>        }
>    }
> 
>    MPI_Buffer_detach(&bsend_buf, &bufsize);
>    free(bsend_buf);
>    MPI_Finalize();
>    return 0;
> }
> 
> 
> 
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss