Newsgroups: comp.parallel.mpi
From: fkrogh@math.jpl.nasa.gov
Subject: Library Software
Organization: Jet Propulsion Laboratory - Pasadena CA
Date: Tue, 15 Oct 96 11:10:33
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Message-ID: <540k2p$4ra@netline-fddi.jpl.nasa.gov>

I am thinking about writing a general linear least squares solver using 
MPI.  It seems to me that there are two environments where such a code 
might be used, and that library software perhaps should be quite 
different for the two environments.  The cases I have in mind are:

1. Supercomputer with homogeneous processors.
2. Network of workstations (which may offer substantially different
   computing power to the application).
   
In the second case I suppose one wants to dynamically move parts of the
matrix from a too busy place to one that is not busy enough.  In the
first case this seems to me liable to be counter productive.

Are there any opinions out there as to whether these two cases should be 
treated differently?  If not, would you advocate dynamically allocating 
parts of the matrix?  If so, which environment in your opinion has the 
greatest need for math library components.  

Thanks for any opinions,
                                Fred
                     
fkrogh@math.jpl.nasa.gov
(818)354-2663                     

