[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Getfem-users] QUterm vs. parallelization
From: |
Yves Renard |
Subject: |
Re: [Getfem-users] QUterm vs. parallelization |
Date: |
Fri, 3 Apr 2009 12:52:12 +0200 |
User-agent: |
KMail/1.9.5 |
Dear Igor,
Unfortunately, the parrallelization of Getfem is still incomplete. You are
right, th brick mdbrick_QU_term is not parallelized. In fact, the
modification to parallelize it is not difficult. You have to replace the last
argument of asm_qu_term, which is boundary, by the term
mf_u.linked_mesh().get_mpi_sub_region(boundary)
in the definition of the brick (getfem_modeling.h).
The effect is that the assembly will be done in a distributed way with respect
to a partition of the mesh done with metis.
In fact, all model brick have to be modified in that way to be used in a
parallel version.
I commit this change in the gna repository.
Best regards,
Yves.
On Friday 03 April 2009 11:29, Igor Peterlik wrote:
> Dear GetFEM team,
>
> I would like to ask about a bit strange behavior of the QUterm brick
> being used with GETFEM_PARA_LEVEL=2.
> First, I noticed that my application composed of the Helmholtz,
> Dirichlet (twice) and QUterm bricks gives
> different results in sequential and parallel version (and moreover the
> result in parallel version depends on the number of
> the processors). I found out that the QUterm brick modifies the entire
> system matrix on each process. Maybe I had missed
> something but it seemed to me that this was not correct behavior
> (maybe QUterm is not compatible with paralel mode...).
> As a workaround, I added a condition to the gmm:add as shown below:
>
> if (getfem::MPI_IS_MASTER())
> gmm::add(get_K(), gmm::sub_matrix(MS.tangent_matrix(), SUBI));
>
> in the mdbrick_QU_term::do_compute_tangent_matrix().
> so the system matrix is modified only once. It seems this helped,
> since the solution is now invariant wrt. the number
> of CPUs. Nevertheless, this is not the real solution I guess.
>
> BTW is there some (at least) small documentation concerning the
> paralelization (which parts are parallelized and how)?
>
> Regards
>
> Igor
>
> _______________________________________________
> Getfem-users mailing list
> address@hidden
> https://mail.gna.org/listinfo/getfem-users
--
Yves Renard (address@hidden) tel : (33) 04.72.43.87.08
Pole de Mathematiques, INSA-Lyon fax : (33) 04.72.43.85.29
20, rue Albert Einstein
69621 Villeurbanne Cedex, FRANCE
http://math.univ-lyon1.fr/~renard
---------