From: salo@cray.com (Eric Salo)
Newsgroups: comp.parallel.mpi
Subject: Re: Blocking wait on fd and MPI ?
Date: 28 Jun 1999 23:27:16 GMT
Organization: Silicon Graphics, Inc.
Message-Id: <7l90ck$46o$1@walter-fddi.cray.com>
References: <slrn7nfp37.6a8l.martial.michel@miyu.cam.nist.gov>
Xref: ukc comp.parallel.mpi:5258


> Is there a simple/hard way to solve this communication/multiplexing
> problem ?

Not really. At least, not without moving outside of the MPI standard.
The best you can hope to do is to roll your own routine which
periodically calls both some flavor of MPI_Test() and select(),
using some sort of backoff algorithm to put the process to sleep
for increasingly long periods of time until something interesting
happens. This will approximate the true blocking behavior that you
are seeking any may or may not be sufficient for your needs.

However, since you are at NIST, I assume that you are a LAM user.
So if you're feeling particularly ambitious, you might try writing
an extended version of MPI_Wait() which also accepts some number
of user-supplied file descriptors and adds those to the internal
select() call used by LAM. But this could be a complicated piece
of work (one of the LAM guys could tell you more) and would have
the disadvantage of being completely non-portable.

--
Eric Salo      Silicon Graphics      salo@sgi.com