Newsgroups: comp.parallel.pvm
From: eve_li@btgmax.zko.dec.com (Eve Li)
Subject: Re: pvm message passing timing
Organization: Digital Equipment Corporation
Date: 18 Aug 1995 09:33:05 GMT
Message-ID: <411moh$bgv@nntpd.lkg.dec.com>

Raja B Daoud (rdaoud@magnus.acs.ohio-state.edu) wrote:
: This is one case where I would take the min, not the average.  Since
: at some time the application was able to achieve this min. timing in
: a shared environment, it's very likely that if you were to take a
: dedicated cluster and average several samples, this "good timing" will
: be closer to the min. than to the average of the shared env. samples.
: To be safe, one can take the average of the few smallest samples (say
: the fastest 10-20% of the shared env. samples) or better yet, plot the
: samples and only consider the cluster of min. samples (i.e. visually
: filtering out the spikes in the graph).

Take minimum sounds like  a good idea. However, an implicit assumption is
that the parallel system performs the same way on two 
systems (loaded and unloaded) given the same set of data.
If the parallel system will do some load balance during the computation, 
and load balance depends on the data processed (thus, same data processed in 
different order will have different effect), the minimum will not be very
meaning in such an environment. 

eve 



: --Raja

: -=-
: Raja Daoud				raja@tbag.osc.edu
: Ohio Supercomputer Center		http://www.osc.edu/lam.html

