From: eugene@cse.ucsc.edu (Eugene Miya)
Newsgroups: comp.parallel
Subject: Re: What's Massive Parallel Processing?
Date: 17 Oct 1998 07:17:07 GMT
Organization: UC Santa Cruz CIS/CE
Approved: eugene
Message-Id: <709g9j$bs1@darkstar.ucsc.edu>
References: <6v6qm2$ckt$1@encore.ece.cmu.edu>


In article <6v6qm2$ckt$1@encore.ece.cmu.edu>,
Jeff Law Boon Hiong  <lbseng@tm.net.my> wrote:
>Can anyone answer me?

Sure many readers could, but why take the time?

I'll try to be true to an introduction Norm Matloff did for me the other day.

MPP: processing done on the Goodyear/Loral MPP.
	Processing attempting to beat the von Neumann bottleneck
	(processor to memory) inefficiencies by replicating large numbers
	of CPUs.
	It's a marketing term.

The trick is "what's large numbers?"
It's a question of scale.

The original MPP had 16K CPUs.  8 per chip  (I was given a reject chip).
But the amount of memory per CPU was small 2Kb and bit serial.  This was
barely acceptable for the kinds of image processing applications
the MPP was built.  ("Ya want floating point with that?")

SO Later the next implementations have first 64K and 32K and down to 16K
and 8 K and 2K processors (The Thinking Machines Corp. Connection Machine
series (CM-1, 2, 200, 5).  And there was also Marpar (barely still is).
Some amusing market charts were created during this time explaining the
lower bounds for MP.  These were bit serial CPUs (until the 5).
And many grand claims were made about the capabilities of these
machines.  But that's history.  Oh, 8 Kb memory per CPU.
("Ya want floating point with that?" that's the 2K version of counting)

Meanwhile other machines were attempting to reach up and claim 256 processors,
and hypercubes with 1024 and 4096 conventional u-processors were
reaching into massive parallelism range.  The hypercube guys attempted
to jump on that bandwagon (that's about the time Steve started this news
group comp.hypercube).
("Ya want fries, I mean floating point, with that?")

There were other players, too: ICL DAP, Meiko's computing surface, 
H-P/Convex, etc.  All very minor, they didn't try the marketing claims
of massive parallelism until after the above got established.

Academically, many people proposed millions of processor elements
(myriaprocessors), but those were all vapor.

There's also the concepts of Fetch-Op and processors in memory work.
Holland get credit for ideas on cellular automata, neural nets, etc.
Limited implementation on all of these.
Some specialized machine with Ks of processors have been built, but not
markets (custom jobs).  GF-11, special image and single processing
machines, etc.  Their numbers, but no claims, fall into the above ranges
of processor numbers.



Non-technical References:

%A George Gilder
%T Hillis vs. Bell: the law of the microcosm
%J Upside
%V ?
%N ?
%D January 1992
%P 24-42
%X Non-technical but interesting characterization (you don't have to agree
with some of the simplifications) of the Hillis-Bell bet on performance
and market share of distributed memory massive-parallel machines versus
shared-memory machines by the end of 1995.
Sides with Bell.
%X Update: January 1996 issue: Hillis pays off Bell (Bell 'wins').

%A Willie Schatz
%T Laying a Bet To Rest
%J Upside
%V ?
%N ?
%D January 1996
%P 39-45
%K massive parallelism,
%O http://www.upside.com/resource/print/9601/teraflop.html
%X See 1992 Upside article on the Gordon Bell-Danny Hillis bet.
George Gilder, "Hillis vs. Bell: the law of the microcosm",
Upside, January 1992, pp. 24-42
%X Gordon Bell claims to be the winner of his bet with Danny Hillis.
The author believes that the real issue isn't who has the fast
machine, but whether anybody will still make Big Iron machines in the
future.  The author believes specifically that there will be a few
niche markets for this kind of hardware, but that general purpose
machines with lots of processors are dead.  Indeed, the author
believes that the market never really happened and that it was really
fed by Cold War paranoia.  The facing page of the cover is a drawing
showing Danny and someone else in a graveyard, with two gravestones
reading RIP Thinking Machines 1983-94 and Kendall Square Research
1986-95, with Danny rolling a pair of dice.  Sprinkled through the
piece is a collection of gravestones with all of the big supercomputer vendors.
The author doesn't particularly side with either Hillis or Bell.
There's a sidebar about Tera Computer and its recent IPO entitled
"$40 Million Vaporware, Anyone? - How's this for a hot prospect?"
%X See also "Communications of the ACM," vol. 39, no. 3, March 1996,
pages 11-15.


Technical refs: in the usual place.

--
Articles to bigrigg+parallel@cs.cmu.edu (Admin: bigrigg@cs.cmu.edu)
Archive: http://www.hensa.ac.uk/parallel/internet/usenet/comp.parallel