We extended LAM to include these datatypes (INTEGER1|2|4|8), but we
didn't modify the reduction algorithms accept the new datatypes
(recally that only certain datatypes are valid with certain reduction
operations). Sorry. :-\
Open MPI already supports all these datatypes and all the relevant
reduction operations on them, so there's little chance that we'll be
extending LAM's reduction operations. Can you hang on for a stable
release of Open MPI in a few months? (the first stable release of Open
MPI is projected for 1Q 2005).
On Nov 12, 2004, at 2:28 PM, Grismer Matthew J Civ AFRL/VAAC wrote:
> There seems to be a problem with the newly included Fortran datatypes,
> at least with MPI_INTEGER4 and MPI_REAL4. Our code kept aborting at
> MPI_Reduce or MPI_Allreduce calls that had one of these as the sending
> datatype:
>
> MPI_Reduce: invalid operation: Invalid argument (rank 0,
> MPI_COMM_WORLD)
>
> Rank (0, MPI_COMM_WORLD): Call stack within LAM:
>
> Rank (0, MPI_COMM_WORLD): - MPI_Reduce()
>
> Rank (0, MPI_COMM_WORLD): - main()
>
> When I changed the datatypes to MPI_INTEGER or MPI_REAL, everything
> was fine. This is a well-tested code with many different
> machines/compilers/MPI libraries, so I'm confident it is not a bug in
> the code. My quick fix was to modify mpif.h so that MPI_INTEGER4 had
> the same value as MPI_INTEGER, MPI_REAL4 had the same value as
> MPI_REAL, and MPI_REAL8 had the same value as MPI_DOUBLE_PRECISION.
> Actually, this is how I had always modified mpif.h in older versions
> of LAM anyway.
>
> Matt
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
