Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!linus!decvax!harpo!eagle!mhuxt!mhuxi!mhuxa!houxm!ihnp4!ixn5c!inuxc!pur-ee!uiucdcs!uicsl!pollack
From: pollack@uicsl.UUCP
Newsgroups: net.ai
Subject: Re: data flow computers and PS's - (nf)
Message-ID: <2573@uiucdcs.UUCP>
Date: Mon, 15-Aug-83 22:44:05 EDT
Article-I.D.: uiucdcs.2573
Posted: Mon Aug 15 22:44:05 1983
Date-Received: Tue, 16-Aug-83 21:27:59 EDT
Lines: 23

#R:sri-arpa:-360300:uicsl:15500002:000:954
uicsl!pollack    Aug 15 19:27:00 1983

The nodes in a data-flow machine, in order to compute efficiently, must
be able to do a local computation.  This is why arithmetic or logical
operations are O.K. to distribute.  Your scheme, however, seems to
require that the database of propositions be available to each node, so
that the known facts can be deduced "instantaneously". This would cause
severe problems with the whole idea of concurrency, because either the
database would have to be replicated and passed through the network, or
an elaborate system of memory locks would need to be established.

The Hearsay system from CMU was one of the early PS's with claims to
a concurrent implementation. There is a paper I remember in IEEE ToC (75
or 76) which discussed the problems of speedup and locks.

Also, I think John Holland (of Michigan?) is currently working on a
parallel PS machine (but doesn't call it that!)


Jordan Pollack
University of Illinois
...!pur-ee!uiucdcs!uicsl!pollack