Newsgroups: comp.parallel.mpi
From: Lawrence Chen <lawrence@combdyn.com>
Subject: Behaviour of MPI/P4 on uniprocessors....
Organization: Combustion Dynamics Ltd
Date: Tue, 11 Mar 1997 16:57:56 -0700
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <3325F184.FB7@combdyn.com>

I'm experimenting with MPI....I have the latest version of MPICH
compiled on
our Solaris x86 2.5.1, HPUX 10.20, and DEC Alpha machine with the p4
device.

I was experimenting with the Poisson's example program from "Using MPI"
(Intermediate).

Trying it as 1, 2, 4, 8 processes, the time to completion seems to grow
in
relation to this.

This is as both some processes over different machines in our office and
all
the processes on the same machine.  The fastest of all cases is a single
process on the Alpha.

IE: np = 1 Alpha = 0.6 secs.
    np = 2 Alpha = 1.1 secs.
    np = 4 Alpha = 2.5 secs.
    np = 8 Alpha = 5.4 secs.

Trying the shmem device is....different.

IE: np = 1 Alpha = 0.2 secs.
    np = 2 Alpha = 2.3 secs.
    np = 4 Alpha = 8.3 secs.
    np = 8 Alpha = 20.7 secs.
    np = 16 Alpha = 195.9 secs.

I'm guessing that the increase in time is the added overhead of the
message
passing and the core memory requirements.

Though in the p4 case the time increase is quite directly related to the
number of processes.

Is this expected behaviour?  Will, I be better to run on a
multiprocessor
Alpha to see the advantage of parallel processing?  Any ideas what kind
of
results I might see?  (would like to know before I get a multiprocessor
machine
to run on....that way at least I won't be surprised when the boss sees
the
results 8-)

-- 
 Lawrence "The Dreamer" Chen, P.Eng.               Computer/Research
Engineer
 Combustion Dynamics Ltd., #203, 132 4th Ave. S.E., Medicine Hat, AB 
T1A 8B5
 Phone: +1 403 529 2162    Fax: +1 403 529 2162   Email:
lawrence@combdyn.com
 "Just a Crazy Engineer with an Amiga and a Newton MP130"      
VE6LKC/VE6PAQ