Newsgroups: comp.parallel.pvm
From: bjo@fwi.uva.nl (Benno Overeinder)
Subject: Re: pvm message passing timing
Organization: FWI, University of Amsterdam
Date: 13 Aug 1995 16:53:44 +0200
Message-ID: <40l3lo$788@word.fwi.uva.nl>

eve_li@btgmax.zko.dec.com (Eve Li) writes:

>Since networks of workstations has radom distrubations, it will be ideal
>to use some kind of experiment design that will get rid of such
>disturbations for the whole experiment. Has anyone  done measurement of 
>parallel systems in such an environment? Is there some better way to 
>compare two or more  such  systems besides taking samples alternatively
>and taking the avaerage (see below)?  

A nice treatment of one-factor experiments can be found in Raj Jain's
"The Art of Computer Systems Performance Analysis."

Basic idea is to collect observations (iterated over i) from different 
systems (or alternatives, iterated over j).   This gives you a series 
of observations y_ij.

The model used is:
    y_ij = mu + alpha_j + e_ij,
where mu is the mean response, alpha_j is the effect of alternative j, 
and e_ij is the error term.  The effects are computed so that they add
up to zero.  Sum(alpha_j) = 0.

All the terms can be computed and an important part of the analysis is 
the allocation of the variation, e.g., how much of the variation is
due to the effect factors and which part is due to errors.  The larger
the explainable part of the variance, the better.

It results in a Analysis of the Variance (ANOVA), after which you
can use the F-test to determine the confidence (for example 0.95,
or 0.99) in the analysis.


Benno
-- 
 Benno Overeinder			 Computer Systems Department	     
 voice: (+31) 20 525 7536		 University of Amsterdam	    
 fax:   (+31) 20 525 7490		 Kruislaan 403, 1098 SJ  Amsterdam
 e-mail: bjo@fwi.uva.nl		 	 The Netherlands

