Newsgroups: comp.parallel.mpi
From: Marcelo Pasin <pasin@imag.fr>
Subject: Re: Q: Thread-Safe MPI Implementation
Organization: IMAG, Grenoble, France
Date: Wed, 02 Oct 1996 18:50:21 +0200
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <32529D4D.4925100A@imag.fr>

Christian Schlegel (schlegel@faw.uni-ulm.de) wrote:
> I am interested in a thread-safe implementation of MPI (Solaris 2.x/Linux).
> Does anybody know of an implementation or has tried to use e.g. LAM in a
> multithreaded environment ?

Well, after having waited an answer fore some days, I decided to post.
I guess thread-safeness of MPI is a major issue now. No one seems to
offer
it in a clean way. Not even IBM-AIX-4.x which has threads in kernel
level
can't turn IBM-SP/2-MPI with several threads.

In our project, we face daily problems concerning threads and MPI and
we still didn't find a good solution. We call the existing MPI versions
always without blocking in it and always protected by a mutex.
Nevertheless,
most of the OS and MPI combinations doesn't work well. We have only two
good
experiences: {IBM-AIX-3.1 + SP/2-MPI-F} or {Solaris 2.x + MPI-CH}. The
first
is well tested and works poorly due to some bugs in the DCE threads
library.
The second has less threads lib bugs but p4 seems to blow up when we use
big
messages.

We even thought about rewriting a thread-safe MPI subset which would run
our software in a clean way. Probably we will do it soon. We also expect
some
answer from the teams who develop MPI, but they don't seems to have an
answer
yet.

-- 
Marcelo Pasin                      email: Marcelo.Pasin@imag.fr
Institut IMAG - LMC/Apache,          tel: (+33) 76 51 45 74
100 Rue des Mathematiques,           fax: (+33) 76 51 48 46
BP 53, Campus Universitaire, 38041 Grenoble Cedex 9, FRANCE

