Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: [OMPI users] Fortran issues on Windows and 1.5 Trunk version
From: Damien (damien_at_[hidden])
Date: 2010-05-12 00:11:00


Hi all,

Me again (poor Shiqing, I know...). I've been trying to get the MUMPS
solver running on Windows with Open-MPI. I can only use the 1.5 branch
because that has Fortran support on Windows and 1.4.2 doesn't. There's
a couple of things going wrong:

First, calls to MPI_Initialized from Fortran report that MPI isn't
initialised (MUMPS has a MPI_Initialized check). If I call
MPI_Initialized from C or C++, it is initialized. I'm not sure what
this means for MPI calls from Fortran, but it could be the cause of the
second problem, which is: If I bypass the MPI_Initialized check in
MUMPS, I can get the solver to start and run in one process. If I try
and run 2 or more processes, all the processes ramp to 100% CPU in the
first parallel section, and sit there with no progress. If I break in
with the debugger, I can usually land on some MPI_IProbe calls,
presumably looking for receives that don't exist, possibly because the
Fortran MPI environment really isn't initialised. After many debugger
break-ins, I end up in a small group of calls, so it's a loop waiting
for something.

For reference, it was yesterday's 1.5 svn trunk, MUMPS 4.9.2, and Intel
Math libraries, and a 32-bit build. MUMPS is Fortran 90/95 but uses the
F77 MPI interfaces. It does run with MPICH2. I realise that 1.5 is a
dev branch, so it might just be too early for this to work. I'd be
grateful for suggestions though. I can build and test this on Linux if
that would help narrow this down.

Damien