Newsgroups: comp.parallel.pvm
From: "Robert J. Block" <rblock>
Subject: Re: seeking opinions about ScaLAPACK
Organization: University of Illinois at Urbana
Date: 12 Sep 1995 16:17:18 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <434bqe$abv@vixen.cso.uiuc.edu>

mubin@wein02.elektro.uni-wuppertal.de (mubin zhuang) wrote:
>
>Has SCALAPACK (the part of solving linear equations) been used 
>widely in parallel computation, especially on loosely-coupled 
>distributed memory architecture parallel computer systems like 
>workstation clusters?
>
>What is your experience in using SCALAPACK to solve linear 
>equations (efficiency, reliability and convenience for use and 
>learning)?


I've been using scalapack 1.0 with pvm3.3.5 on a network on Sun 
workstations.  It also took me some time to get the data distribution
format right, and using some of the routines can be a bit messy.
For example, routine pdsyevx (eigenvals/vecs) takes no fewer than
29 parameters, and documentation explaining how large to make 
parameter #23 (length of local work array) was over half a page long.

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.)

Reliability - about as reliable as pvm and the state of the machines that 
are hooked up (okay for my purposes, can't say about production environment).

Convenience - like I said, rough around the edges.  The app. I'm building
is actually a dist-memory matlab-type environment, which is designed to
make some of scalapack's technicalities transparent. (It's got a long way
to go.)  In general, I would say writing Fortran apps. using scalapack is 
pretty tedious because of the data distribution req. and user interface.

Just my $.02

--
Robert J. Block      | University of Illinois at Urbana-Champaign
rblock@ncsa.uiuc.edu | <URL: http://www.ncsa.uiuc.edu/People/rblock>


