Path: utzoo!attcan!uunet!ncrlnk!ncrcae!hubcap!wsmith
From: wsmith@dopey.MDBS.UUCP (Bill Smith)
Newsgroups: comp.parallel
Subject: Re: Information on Massively parallel machines
Summary: It isn't necessarily that bad.
Message-ID: <7153@hubcap.clemson.edu>
Date: 21 Nov 89 13:38:22 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 37
Approved: parallel@hubcap.clemson.edu

In article <7070@hubcap.clemson.edu>, ut-emx!gary@cs.utexas.edu (Gary Smith) writes:
> 
> I must admit I remain very skeptical that massive MIMD parallelism will
> pay off in the near future, if ever.  Now modest MIMD parallelism, up to
> perhaps 256 processors might, assuming we get a significant number of
> applications that can achieve parallelism in excess of 99.9%, and if syn-
> chronization overhead can be made arbitrarily small.  Of course, these
> are pretty severe constraints!
> 
> The matrix below is the reason for my skepticism.  I included the short C
> program I used to generate the F's.
	[ matrix eliminated ]


There are several reasons that this conclusion is not valid.

1.   It is not necessary for speedup to be linear for a parallel processor
	to be worthwhile.   All that is wanted is faster performance.
	How much one is willing to pay for that performance dictates when
	one should start spending more money on bigger processors vs.
	when one should get more smaller processors.

2.   Many problems parallelize naturally.  With such problems, it is
	easy to get a highly efficient parallel implementation.
	Examples include computer graphics or complicated numeric processing

3.   Just because one does not get 100% utilization of a processor on
	one problem, it does not imply that computational power will be
	wasted.   Timesharing systems are a perfect example.   No one
	user uses 100% of the computer, but the sum total of the users
	do get nearly that utilization from the computer.   MIMD
	computers, in principle, could be timeshared as well.   Shared
	memory computers definitely are timeshared.

Bill Smith
pur-ee!mdbs!wsmith
(Not necessarily opinions of MDBS Inc.)