Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!gem.mps.ohio-state.edu!samsung!uunet!ncrlnk!ncrcae!hubcap!craig
From: craig@Think.COM (Craig Stanfill)
Newsgroups: comp.parallel
Subject: Re: Information on Massively parallel machines
Message-ID: <7160@hubcap.clemson.edu>
Date: 21 Nov 89 19:19:32 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 16
Approved: parallel@hubcap.clemson.edu

I wish I had time to respond to this in more depth, but let me take a
quick stab at this from an informal point-of-view, with a single
application: 3D aerodynamics.  Let us assume that one wishes to
simulate 3D airflow over an airframe, using a 1024 x 1024 x 1024 grid,
for a total of 2**30 cells.  It is probably reasonable to put, for
example, a 16x32x32 subgrid (2**14 points) on each processing element;
this achieves a good ratio of local to non-local computation.  One
then needs 2**16 processors.  Our experience is that, under these
circumstances, the serial front end is not the bottleneck. 

This is, by the way, a problem that people very much want to solve. 

				Craig Stanfill
				Advanced Information Systems
				Thinking Machines Corporation
				(617) 876-1111