On Jun 10, 2006, at 7:42 PM, Jeff Whitaker wrote:
> I'm trying to run an operational weather prediction code from the
> National Weather Service on a cluster of dual-G5 macs using xlf and
> lam
> 7.1.2. On the first call to mpi_allreduce, I get this error
>
> MPI_Allreduce: invalid operation: Invalid argument (rank 0,
> MPI_COMM_WORLD)
> Rank (0, MPI_COMM_WORLD): Call stack within LAM:
> Rank (0, MPI_COMM_WORLD): - MPI_Allreduce()
> Rank (0, MPI_COMM_WORLD): - main()
>
> where the call looks like this
>
> call mpi_allreduce(super_val,super_val1,superp+1,mpi_rtype,&
> mpi_sum,mpi_comm_world,ierror)
>
>
> mpi_rtype is set to mpi_real8, and super_val and super_val1 are real*8
> arrays of length superp+1. The same code works on a linux cluster
> with
> mpich. Is mpi_allreduce supported with mpi_real8 and mpi_sum?
LAM does not support using the MPI predefined operations with the
optional Fortran datatypes (such as MPI_REAL8). It is unlikely we
will be adding such support, as we are winding down support for LAM/
MPI in favor of Open MPI. We'll gladly accept patching adding such
functionality, of course. If switching to Open MPI is not an option
and hacking at the internals of LAM isn't your thing, the final
option is to provide a user-defined reduction operation that can deal
with MPI_REAL8 datatypes and use that instead of MPI_SUM. Hardly
perfect, but it will work in a pinch.
By the way, just for clarification... The MPI standard is really
fuzzy on whether MPI_SUM is required to work with MPI_REAL8. It's
not explicitly listed in the types MPI_SUM should work with. On the
other hand, it seems pretty dumb to have a predefined datatype that
isn't supported by MPI_SUM. This happened in LAM because MPI_REAL8
was added much later than the other datatypes and whomever added it
simply forgot to update the reduction code.
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
