From: Mark Gebhardt Newsgroups: comp.parallel.mpi Subject: MPI performance worse than PVM??? Date: Mon, 15 Mar 1999 18:16:26 +0200 Organization: RRSG, University of Cape Town Message-Id: <36ED325A.69D8C9B8@rrsg.ee.uct.ac.za> Reply-To: mark@rrsg.ee.uct.ac.za Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.parallel.mpi:4757 Hi all I have come across what appears to be a strange phenomenon in MPICH (version 1.1.2) and wondered if anyone could shed some light: I have an app. which I implemented in PVM and have now ported to MPI for purposes of comparison (and learning). It is a type of genetic algorithm for antenna array design - however, this is not important. Under certain conditions, my MPI implementation gives wildly varying execution times - something that the PVM version never did. Here's what happens: I have 8 machines in my cluster and I have tested my app using different numbers of processors and the same task (hoping that increasing processors will decrease computation time, as this is how the app is designed!). This proves to be true for up to 8 processors (i.e 1 master and 7 slaves, one process per processor). Time for 2 processors = 2.26 seconds, time for 8 processors=.56 seconds - nice! But when I use 9 processes (ie the machine with the master also has a slave). There is considerable slowdown of the whole process. Now this is fine and to be expected, as there are now 2 processes fighting for cpu cycles, but the problem is that the processing time varies tremendously (from .55 seconds to 1.1 seconds). This only seems to happen if I run 2 processes on the machine with the master process (ie 2 slaves on one cpu produces a slowdown, but gives consistent timings) Does anyone know why MPICH is affected so badly by this competition for the cpu? I am especially interested as my PVM app. shows no such problems. Thanks in advance Mark -- Mark Gebhardt Radar Remote Sensing Group Department of Electrical Engineering University of Cape Town Private Bag, Rondebosch South Africa www: http://rrsg.ee.uct.ac.za/