Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: Notesfiles; site lll-crg.UUCP
Path: utzoo!watmath!clyde!bonnie!akgua!sdcsvax!sdcrdcf!hplabs!hpisla!hplvla!lll-crg!brooks
From: brooks@lll-crg.UUCP (brooks)
Newsgroups: net.arch
Subject: Re: Re: Transputer and occam
Message-ID: <489@lll-crg.UUCP>
Date: Mon, 1-Apr-85 16:18:00 EST
Article-I.D.: lll-crg.489
Posted: Mon Apr  1 16:18:00 1985
Date-Received: Sun, 14-Apr-85 03:55:04 EST
References: <675@cmcl2.UUCP>
Organization: Hewlett-Packard
Lines: 26

> > >I'm a little dubious about the value of hypercubes, as most big
> > >programs have a 5% to 20% purely serial component.
> > >(Note: this only applies to general-purpose
> > >machines.  Obviously, certain problems can have the serial part reduced
> > >to a tiny fraction.)
> > 
> > Have you any data to support this claim?
> 
> There is one very large piece of supporting data:  CDC's incredibly
> expensive failure, the Star-100 supercomputer, which ran vectors very
> quickly and scalars very slowly.  They had hoped for automatic vectorizers
> that would be good enough to make the thing work, and had also hoped that
> most problems would vectorize simply and almost completely.  No such luck.
> One of the reasons the Cray-1 was such a success, by comparison, was that
> the Cray does scalar operations at blazing speed too.
> -- 
> 				Henry Spencer @ U of Toronto Zoology
> 				{allegra,ihnp4,linus,decvax}!utzoo!henry

One should be careful about claiming that the failure a machine with a high
penalty for scalar processing vs vector processing transfers to MIMD machines
in general.  Vectorization is a form of parallelism but is a rather restricted
and highly regular form.  An application may have parallelism that can't be
exploited with vectorization.  The most attractive parallel processor
architectures allow the exploitation of both forms.  The idea is to have many
cpus sharing memory and give each cpu has vector instructions.