OpenMPI bug?

Hi JMB

actually there are 2 versions of Aster in CAELinux 2010:
- one that is part of Salome-Meca which was compiled by EDF using Intel compiler with OpenMP support (nbcpu>1) for shared memory parallelism on the standard solver Mult_Front. But this version does not support MPI Mumps or PETSC solvers.
This solver is installed in /opt/SALOME-MECA*/aster.
To use it from terminal, you will need to source the Salome-Meca environment:
[code:1]. /opt/SALOME-MECA*/envSalomeMeca.sh[/code:1]
( or something similar, I don't remember the exact path by heart)

-the second was compiled with GNU compilers and ACML libs without OpenMP support (nbcpu must be =1, no paralelism of Mult_front) but with MPI support for Mumps and Petsc solvers (mpi_nbcpu>=1).
This version is installed in /opt/aster101 and is the one that is available from Astk when you run it from the CAElinux menu.
The reason for the lack of openmp support is because of a bug of Gnu compiler when handling openmp shared arrays... so not easily fixable.

To make it short :
if you want to use a parallel Mult-front solver, you will need OpenMP so you must use the version from Salome-Meca
if you want to use a parallel Mumps or petsc solver, you need MPI and thus the version that I compiled

There is no reason to use both nbcpu>1 and mpi_nbcpu>1 or mpi_nbnoeud >1 because no solver benefits from both OpenMP and MPI.

I hope that it clarifies this complicated subject..

Hi,

normally the default solver should not be changed, so you should not see such improvement in calculation speed wit mult-front... strange.
Maybe the assembly phase is carried out in parallel ?

Another thing that I forgot to mention, to use MUMPS you should select RENUM = PORD as I did not manage to compile METIS for theparallel MUMPS solver.. it is probably not as optimal but it works.

Joel