Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84; site watrose.UUCP
Path: utzoo!watmath!watnot!watrose!cdshaw
From: cdshaw@watrose.UUCP (Chris Shaw)
Newsgroups: net.arch
Subject: Re: Current Status of Karp Challenge
Message-ID: <8108@watrose.UUCP>
Date: Sun, 27-Jul-86 23:43:21 EDT
Article-I.D.: watrose.8108
Posted: Sun Jul 27 23:43:21 1986
Date-Received: Mon, 28-Jul-86 03:58:10 EDT
References: <1578@ames.UUCP>
Reply-To: cdshaw@watrose.UUCP (Chris Shaw)
Distribution: net
Organization: U of Waterloo, Ontario
Lines: 38
Summary: Pretty easy

In article <1578@ames.UUCP> eugene@ames.UUCP (Eugene Miya) writes:
>First, there is some misunderstanding about the challenge.  The challenge
>was to demonstrate some factor of 100 simple speedup of some program
>running on a computer.  So, the Connection Machine with 64K processors
>falls down in two areas: 1) the computer must be an MIMD rather than
>SIMD type (Multiple Instruction Stream Multiple Data Stream from
>Flynn's [1966] classification scheme).  2) The program must be capable of
>running on a single CPU of said computer.
>
>Second, the only person to even suggest a possible challenge had an ICL DAP.
>
>--eugene miya
>  NASA Ames Research Center

Actually, someone with a 128-node hypercube machine could probably do this.
Finding eigenvalues from a tridiagonal matrix using Bisection is very parallel.
Cleve Moler has a routine which does this. We have run it here at Waterloo on 
our 64-node Intel hypercube, and the speedup attained by simply throwing 
processors at the problem is remarkable. Unfortunately, I can't remember the
numbers.

Also, a local user has a Ray Tracing program which is almost entirely without 
communication, and could therefore run above the required speedup number.
A local copy of the scene is kept on each node, and each processor is 
responsible for a scan line. The results are then sent back, and a new scan 
line is assigned. Given a 128-node hypercube, I would fully expect 126-127
or so speedup. 

Both applications fit the two requirements: The iPSC is MIMD, and they can both 
run on 1 node.

So. Anyone got a spare iPSC-d6 (64 nodes) they could lend us?

Also, since I can name two, would we get $200 ??   B-)

Chris Shaw    watmath!watcal!cdshaw  or  cdshaw@watmath
University of Waterloo
Bogus as HELL!!