From: John Galambos <jdg@ornl.gov>
Newsgroups: comp.parallel.pvm
Subject: Re: a design question
Date: Fri, 16 Jul 1999 14:54:35 -0400
Organization: ORNL
Message-Id: <378F7FEB.8731EC6B@ornl.gov>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Xref: ukc comp.parallel.pvm:8570


"J. E. Lewis" wrote:

> I would like to get an opinion for the design I'm using for a project using PVM.
>
> Here is the problem setup  I have a large quantity of numbers to calculate, and
> each number is associated with certain circumstances (so I need to keep the association)
>
> I am currently using a master/slave design where the master delegates to the slaves to
> calculate the number.  The question is would it be better to spawn enough processes to
> calculate all of them at once or should I only spawn a subset of the necessary number?
>
> There are approximately 100 items that need their number calculated, and I only have 7 or
> 8 machines to use.

I once used pvm for for a genetic algorithm optimization problem (seems similar to the
problem you describe). I spawned only one job at a time per machine. This made recovery
easier if a machine crashed. Also, my "function evaluator processors" that were running in
parallel were pretty big, and used too much memory to put a bunch at once on a single node.
--

John Galambos
Voice: 423-576-5482,  Fax: 423-576-5436
SNS Building, ORNL, Oak Ridge Tn., 37831-8218