Newsgroups: comp.parallel.pvm
From: eddemain@neumann.uwaterloo.ca (Erik Demaine)
Subject: Re: seeking opinions about ScaLAPACK
Organization: University of Waterloo
Date: Tue, 12 Sep 1995 20:44:10 GMT
Message-ID: <DEt89M.Gw5@undergrad.math.uwaterloo.ca>

Robert J. Block (rblock) wrote:
: ...
: Efficiency - well, solving linear systems in this platform usually takes 
: longer in parallel, because of the comm. overhead (I wasn't really expecting 
: any efficiency here, this is only a testing/debugging environment.)
: ...

What sort of platform are you using?  I've been writing some parallel
Cholesky linear solvers for banded matrices in MPI.  Since the band is
around 86, the messages (which consist of a single column each) are only
around 200 bytes.  This leads to a large overhead, and so when I run e.g.
the fan-in algorithm in parallel ON AN SP-2, I get a superlinear speed-down.

Are there any algorithms which use larger messages?  I've looked at ScaLAPACK,
but there seems to be no support for sparse matrices.

The SP-2 has a 40 gigabyte/sec network, however, when using message sizes
of length 200 or so bytes, the speed reduces by a factor of 100.

Does anyone have any ideas on how to get some speedup?? :)

Erik
--
Erik Demaine                 \/  Masters student in Computer Science
Dept. of Computer Science    \/  e-mail: eddemain@neumann.uwaterloo.ca
University of Waterloo       \/  URL   : http://barrow.uwaterloo.ca/~eddemain/
Waterloo, ON Canada N2L 3G1  \/  <<< Space for sale >>>

