Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uwm.edu!gem.mps.ohio-state.edu!tut.cis.ohio-state.edu!purdue!bu-cs!xylogics!world!bzs
From: bzs@world.std.com (Barry Shein)
Newsgroups: comp.arch
Subject: Re: parallel systems
Message-ID: <1989Oct21.234930.905@world.std.com>
Date: 21 Oct 89 23:49:30 GMT
References: <10164@encore.Encore.COM> <1989Oct20.165407.5887@mentor.com>
Organization: The World
Lines: 61


Why oh why is are people searching around for a philosopher's stone of
computing architectures. This isn't science, this is aberrant
psychology.

There exists problems which are best run on (pick one or more) MIMD,
SIMD, scalar, vector-processors, and/or hybrids. There also exist
problems which don't care what they're run on, at least not much.

For example, time-sharing and typical database environments seem to
run very well on MIMD systems for very little re-programming effort
(in the case of time-sharing, none.)  This is because of the large
granularity of the applications and their measuring of "runs well" as
mere response time.

There are other examples for MIMD and examples for SIMD and other
types of processors.

To say (MIMD/SIMD) processors are a bad idea because there exists some
large set of problems which are either impossible or very hard to
optimize for those architectures is so goddamn stupid it boggles the
mind.

MIMD processors are relatively easy and cheap to build out of mostly
commodity parts.

SIMD processors appear to be very, very good at certain classes of
problems which are very important to some people. Important enough
that they'll buy a SIMD box just to run those problems and tell people
with other problems that don't fit so well to go fly a kite (which is
the mathematically correct answer to them.)

We've had multi-processing and hardware optimizations almost since
computing began. What do you think makes a mainframe a mainframe?
Multi-processing, particularly in the I/O channels because mainframes
are bought for high I/O throughput. Most DP shops aren't CPU bound,
they're I/O bound, so they buy their CPUs in the channels.

It continues to astounds me how, particularly in the academic computer
science community, some dodo will stand in front of an audience, show
that there exists a class of problems which don't run well on parallel
architectures, and conclude that therefore parallel architectures are
bad.

What *frightens* me is that N people will sit in the audience and nod
their heads in agreement and go out to spread this gospel (as we've
seen on this list) instead of riding the dodo out on the first rail.

LOOK, there are folks out there using all these architectures and
winning big. Consider that before you attempt to prove on paper again
that it's impossible for a honeybee to fly.

What would be *useful* would be a taxonomy of algorithms classified by
how well (and difficulty of adaptation) to various architectures.

But it's so much easier to throw peanut shells from the bleachers.
-- 
        -Barry Shein

Software Tool & Die, Purveyors to the Trade         | bzs@world.std.com
1330 Beacon St, Brookline, MA 02146, (617) 739-0202 | {xylogics,uunet}world!bzs