Newsgroups: comp.parallel.mpi
From: Armando Lins Netto <armando@me.berkeley.edu>
Reply-To: armando@me.berkeley.edu
Subject: lam mpi error with linux
Organization: U.C. Berkeley
Date: Fri, 24 Jul 1998 15:09:29 -0700
Mime-Version: 1.0
Content-Type: text/plain; charset=iso-8859-2; x-mac-type="54455854"; x-mac-creator="4D4F5353"
Content-Transfer-Encoding: 7bit
Message-ID: <35B9060A.FBAE5E37@me.berkeley.edu>

I am running lam mpi in a linux system and I've been getting an error
after the code runs for a while (hours up to a day).

Basically what happens is that one or more messages get lost for no
aparent reason. The messages cannot be found with mpimsg and mpitask
does not responde after this.

If I run the code again with exactly the same parameters (it should give
the same results in my code) and the problems does not repeat itself in
the same point of the simulation. It happens again in a different
location......

When I run a debbuger attached to all processes, I found out that a
segmentation fault error occurs in a mpilam routine (lam_send () and
others) and crashes everything. All the parameters for the MPI_Send were
within proper range and correc as far as what my code expects. The
interesting thing is that only processes in the same machine get this
segmentation fault. When I run again it can happen at other machine but
then it only affects this new machines's processes.

I have reduced the number of pending messages in the receive side and
the program run for more time than before, but it still crashes......

Does anyone know what can be happening and what to do to fix it?

Thanks for any help!
Armando