Newsgroups: comp.parallel.pvm,comp.parallel
From: Stefan Hinker <hinker@informatik.uni-wuerzburg.de>
Subject: Re: [Q]:Memory Access and Efficiency of parallel (PVM) programs
Organization: University of Wuerzburg, Germany
Date: Thu, 2 Nov 1995 10:21:53 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <47a641$j94@winx03.informatik.uni-wuerzburg.de>

Antonio Scotti <scotti@centauro.upc.es> wrote:
>Which of the following is more efficient and why?:
>
>1) Having the master reading the matrix(es) and multicast it(them)
>   to all the slaves.
>
>2)Having each slave to access the memory to retrieve the data it needs? 

IMHO, it depends on how your hosts are connected.  
a. via ethernet and nfs, on one filesystem:
    both ways would expect to preform about as good, with a tendancy to
    prefering option 1 for increasing numbers of clients and growing datasize.
 
    Reason: ethernet prefers few, big packets
            using nfs blocks your network and needs the ethernet...

b.  via ethernet, not on the same filesystems:
     option 1 seems the only one practical ;-) 

c.  on a multiprocessing-machine with shared memory:
     The MP-pvmd preforms very well, using shared memory anyway.  So why bother
     inventing the wheel all over?

I'm looking forward to other suggestions!

Greetings,
Stefan
-- 
===============================================================================
Stefan Hinker                         email: hinker@informatik.uni-wuerzburg.de
                Unix? Still working on it  :-))
===============================================================================


