Newsgroups: comp.parallel.mpi
From: "Andrew Mc.Ghee" <mcghee@sun.mech.uq.edu.au>
Subject: Non-blocking communication and MPI_Test
Organization: Mechanical Engineering, UQ
Date: Sun, 29 Dec 1996 18:54:28 +1000
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <32C62F49.702C@sun.mech.uq.edu.au>

I'm curious to know if anyone can re-assure me that for non-blocking
communication in parallel code, the MPI_Test routine is all that is
required to free up the communication tags.
Our standard communication routine looks something like this
MPI_Irecv ...
MPI_Isend ...
MPI_Test (for recv)
MPI_Test (for send, to clean up send tags)

This appears to work fine on the SP2 machines, but I appear to be having
some problems on the VPP300 machines useing MPI. It could possible be a
coding problem (The IBM is a wonderfully forgiving machine in terms of
compilers), but I'd like to aleviate a concern that I have that might be
that MPI_Wait is the only way to de-allocate buffer space and tags used
for message passing, and that an MPI_Test is insufficient.
(The problem we are seeing on the VPP is that after a certain amount of
time, an MPI_Irecv is never being full-filled, a problem we once saw on
the SP2 using both MPC and MPI when message tags were not freed up)

Any advice would be greatfully accepted,
regards,
Andrew Mc.Ghee
Mechanical Engineering, University of Queensland
mcghee@sun.mech.uq.edu.au

