Newsgroups: comp.parallel.pvm
From: mubin@wein02.elektro.uni-wuppertal.de (mubin zhuang)
Subject: seeking opinions about ScaLAPACK
Organization: University of Wuppertal
Date: 12 Sep 1995 08:08:42 GMT
Message-ID: <433f6a$oap@wmwap1.math.Uni-Wuppertal.DE>



Dear colleagues,

I am seeking opinions about SCALAPACK (the part of solving 
linear equations). I appreciate it very much if the following 
subject can be discussed in the newsreader. You may of course 
send opinions directly to me if you do not want to open your 
opinions due to some reasons.


===================================

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)?

===================================


My comments (as a computer engineer):

The topic of my current research is parallelism of large-scale
circuit simulation. The key point is parallelization of solving
linear equations.

I used SCALAPACK1.0 (the version released last year) for about a 
half year. The computer resource is a loosely-coupled distributed 
memory architecture parallel computer system: a workstation cluster 
of 3 DECstation-5000. The workstations are connected through Ethernet 
and I installed PVM3.3.7 for parallel computing.

SCALAPACK is a huge software package with numerous and complex interfaces 
of routines.  I found it is not very easy to learn and use. For example, 
there is a lack of  necessary  guides for  distributing the coefficient 
matrix of linear equations into a 2-D block cyclic format. Nevertheless, 
the distribution is  required  when  using  SCALAPACK  to  solve linear 
equations.  So, at that time,  I would rather write a program by myself
with PVM3.3 and F77. My program is very simple and very easy to use, and 
it has high efficiency and reliability working on the workstation cluster. 
It is however a pity that I have not compared  efficiency of my program 
with the new version of SCALAPACK. I think the efficiency of my program 
should  be not  worse  than  SCALAPACK (LU and Solver)  according to my 
experience using the a previous version of SCALAPACK.

I know "software-engineering" and I understand that we may be benefited 
from using the standard software packages when we develop a new software. 
But I think that it maybe more efficient to develop parallel algorithms 
which are suitable to new computer architecture. Does SCALAPACK inherit 
too much from its predecessor e.g. LAPACK etc.?

On the other hand, I heard opinions that there is almost nothing to do 
concerning  development  of parallel  algorithms  for solving linear 
equations due to emergence of SCALAPACK.

What are your opinions?


Mubin

e-mail address:

mubin@wein02.elektro.uni-wuppertal.de







