Newsgroups: comp.parallel.pvm
From: Darren Porras <dporras@crml.uab.edu>
Subject: PVM PERFORMANCE
Organization: University of Alabama at Birmingham
Date: Wed, 25 Jun 1997 18:03:45 -0500
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <33B1A3D1.6BA3@crml.uab.edu>

Does anyone know of ways to reduce PVM communication overhead? I have
tried various optimizations (such as PVMDIRECTROUTE and PVMDATAINPLACE)
but have not observed any real decrease in wall clock times of my
application (simulation). My master process runs on one workstation and
spawns slaves on three other workstations. At every time step data is
sent to each of the slaves and the slaves use this for computation. The
result from each slave is passed back to the master at every time step.
Is there anything I could do to reduce communication overhead? Is MPI
faster than PVM for parallelization across a homogeneous cluster of 
workstations?

Another general question I have is what is the best way to measure
communication overhead. I've been using wall clock time of my 
application as a guide but is this the only way? 


Thanks!

Darren Porras