Newsgroups: comp.parallel.pvm
From: jeff@Adeno.wistar.upenn.edu (Jeff Taylor)
Subject: Re: perf diff b/w spawing on single vs many machines
Organization: University of Pennsylvania
Date: 1 Mar 1996 19:30:46 GMT
Message-ID: <4h7j96$rdg@netnews.upenn.edu>

In article <31351B15.15FB@ece.utexas.edu>,
	Saghir Ahmed Shaikh <shaikh@ece.utexas.edu> writes:
>I have developed a master slave simulator on Nwtwork od stations. using
>pvm.
>
>I wanted to measure the perf. 
>
>It turned out  to be that if I spawn same no. of process on one
>workstations, it is faster as compared to if i spawn the same no. of
>processes in let's say 4-5 workstations. Now I wonder that is it beacuse
>
>(i) data format changes
>(ii) Or pvm deamon knows it that I have to sent a message to the same
>machine so it takes less no. of time? 
>
>
>Please suggest what is going on?
>
>thanks my email is saghir@mail.utexas.edu
>

Saghir -
	This has to be a granularity problem.  In other words, each child
process you spawn is not doing enough work to overcome the overhead that 
comes with spawning the child process and the network delays in communicating 
with it.  Try using the Xpvm program to trace the execution of your program.
This should give you a better idea of what's happening.  Also I've written
a short paper which you can get via anonymous ftp from :
	adeno.wistar.upenn.edu/pub/ga


=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Jeffrey S. Taylor				Protein Crystallography
The Wistar Institute				Molecular Modelling
University of Pennsylvania			Software development
Philadelphia, PA 19104				Molecular Biology
(215) 898-2202			E-Mail    Jeff@aries.wistar.upenn.edu	
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
delays.


