home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: sci.math.symbolic
- Path: sparky!uunet!mcsun!dxcern!news
- From: vermaser@cern.ch (Josef Vermaseren)
- Subject: Re: MPP implementations of symbolic algebra systems
- Message-ID: <1992Sep5.150031.14136@dxcern.cern.ch>
- Sender: news@dxcern.cern.ch (USENET News System)
- Organization: CERN European Lab for Particle Physics
- References: <5!lngnc@lynx.unm.edu>
- Date: Sat, 5 Sep 1992 15:00:31 GMT
- Lines: 30
-
- In article <5!lngnc@lynx.unm.edu> john@aquarius.unm.edu (John Prentice) writes:
- > Anyone know if people are working on implementations of symbolic algebra
- > systems on massively parallel supercomputers? If so, what precisely is
- > being parallelized?
- >
- I have been experimenting with putting FORM on the ACP/MAPS machine at
- Fermilab. At the time I did this the machine had 258 processors. We got
- it to run some moderately complex problem at a speed of more than 100 times
- the speed of a single processor. The fun part was that
- 1: the whole program FORM parallellizes.
- 2: there is only one bottleneck.
- 3: the more complicated the problem, the better it works.
- We ran a simple program and then it could keep only 6 processors going,
- because by the time it was dishing out the 7-th step in a giant tree
- the first processor would be available again.
- The special property of this machine is that the processors can get at each
- others memory rather quickly, so transferring part of a problem to another
- processor is rather fast. So expanding trees can be spread out very quickly,
- provided that the 'environment' that comes along doesn't contain too many
- pointers. With FORM there are hardly any. The crunch comes of course when
- all the results have to be put together. That is a variation on a sorting
- algorithm and actually it is an inverse tree. So most effort should go into
- making the last stap in that inverse tree as insignificant as possible and
- definitely as fast as possible.
-
- The project isn't quite finished yet, and will probably be continued some time
- next year. By now the machine has even more and faster processors so it should
- be fun.
-
- Jos Vermaseren
-