Newsgroups: comp.parallel.mpi
From: "Steve T. Barnard" <barnard@nas.nasa.gov>
Subject: Re: MPI_REQUEST_MAX
Organization: NAS - NASA Ames Research Center, Moffett Field, CA
Date: Thu, 28 May 1998 13:04:51 -0700
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <356DC363.95C79DE9@nas.nasa.gov>

James Cownie wrote:
> 
> Steve T. Barnard wrote:
> >
> > My program is failing with the error:
> >
> > Insufficient request entries.
> > Please use MPI_REQUEST_MAX to increase your allocation.
> > The current value of MPI_REQUEST_MAX is 16384
> >
> > This is on an SGI Origin2000. I've increased the MPI_REQUEST_MAX
> > environment variable to 65000, but that has no effect. I've found that
> > 16384 is the maximum allowable.
> >
> > In any case, I don't consider that to be a proper fix, even if it would
> > work. I'd like to put checks in my program so I never exceed the limit,
> > but I don't understand what the nature of the error is. Can anyone
> > explain it, and suggest how I can avoid it?
> >
> >         Steve Barnard
> 
> Sounds like you're starting a load of requests (e.g. by doing
> MPI_Isend, MPI_Irecv), but never releasing the request objects
> again by doing an MPI_Wait or successful MPI_Test.
> 
> Of course, it could also be a bug in SGI's MPI...

It appears to be a bug in SGI's MPI. All my Isends and Irecvs are
matched with MPI_Test or MPI_Testall. This code works on the SP2, the
T3E, and various workstation clusters. Setting MPI_REQUEST_MAX to 65000
used to fix the problem on the Origin, but no longer.

	Steve Barnard