Path: utzoo!mnetor!uunet!husc6!bbn!gatech!hubcap!peskin
From: peskin@caip.rutgers.edu (R. L. Peskin)
Newsgroups: comp.parallel
Subject: Sandia scaling results
Message-ID: <1224@hubcap.UUCP>
Date: 28 Mar 88 13:23:17 GMT
Sender: fpst@hubcap.UUCP
Lines: 42
Keywords: speedup
Approved: parallel@hubcap.clemson.edu

The recently announced results from Sandia are very impressive, but I
am confused. Part of my problem stems from having to read about this
significant advance in the NY Times rather than from a more
technically informative source. The Times only told of the claimed
speedup, the winning of the "prize", and something about howq they
made the problem bigger rather than smaller. No mention of what
problem was solved. 

Linear speed up with number of processors is not a new result. Well
over a year ago we showed an ABC flow (Beltrami fluid flow) result
that exhibited speedup of 99% of 128 running on a 128 node NCUBE. The
problem parallelized well (obviously) and actually ran at 1.75 times
faster than the same problem on a single processor Cray XMP. Similar
results were reported by U. of Michigan (if my memory serves me) also
on an NCUBE. So there has to more to the Sandia result than just
speedup proportional to the number of processors (they have a 1024
node NCUBE). What is it?

Are there results application domain independant? That would be a real
advance, but improbable. Parallel computing (on distributed systems,
at least) is much like the old analog computing. The name of the game
there was reconfiguring the machine (using a patch-board) to best
represent the physical aspect of the problem. In parallel computing we
have a similar need to decompose the problem domain by appropriate use
of interprocessor messaging and localized computation. I feel this is
very problem domain specific. (It is also a difficult area requiring
knowledge of the application as well as computing per se.) Real gains
will come as general principles evolve that point to methods to
optimize the many choices in domain decomposition. 

Has Sandia come up with such principles? If so, will they let the rest
of us in on it? The "press" has been overwhelming on this one (I even
saw it reported in a NJ "shopping" weekly.) But the technical
community (who were not able to attend the meeting where this was
reported) seem to be in the dark. (I will admit to hearing about this
directly from John Gustafson during a phone conversation, but there
was not time to explore it fully.)

How about it Sandia; can you let us know some facts via a message to
this bboard?

--dick peskin