Newsgroups: comp.parallel.mpi
From: lusk@donner.mcs.anl.gov (Rusty Lusk)
Subject: Re: Size Problems with MPICH
Organization: Argonne National Laboratory
Date: 15 Oct 1996 18:41:05 GMT
Message-ID: <540ls1$27n@milo.mcs.anl.gov>

In article <3263C006.2CB7@lirmm.fr>, "Y. Hamadi" <hamadi@lirmm.fr> writes:
|> Hi, 
|> I've the following problem when using MPICH on a network of five Sparc5
|> all running under Solaris 2.5
|> 
|> mpirun -v -np 40 a.out 40 3 1 .5
|> running a.out on 40 solaris ch_p4 processors
|> accept: Too many open files
|> 
|> What is the solution, i need a lot of process, 40 process is a small
|> value for my application!!!

The MPI model more or less assumes that there will be one CPU per MPI
process.  That is, you can reasonably expect efficiency to drop severely
if you multiplex your MPI application processes onto a smaller number
of CPU's.  Another way to put this is that the arrangement of the application's
"virtual processes" onto MPI processes is the users responsibility.

The particular error you are seeing can probably be avoided by reconfiguring
your system to allow more open sockets at one time (MPICH does not route
messages through intermediate daemons), but that is not the best solution.
A better approach would be to arrange your algorithm so that regardless of the
number of tasks you have to do, the number of MPI processes created matches
the number of CPU's that you have to run them on.

Regards,
Rusty

