Xref: utzoo comp.sys.ncr:521 comp.arch:18373
Path: utzoo!attcan!uunet!wuarchive!uwm.edu!ux1.cso.uiuc.edu!dino!news.iastate.edu!du248-02.cc.iastate.edu!carter
From: carter@du248-02.cc.iastate.edu (Carter Michael Brannon)
Newsgroups: comp.sys.ncr,comp.arch
Subject: Re: Terradata architecures
Keywords: YNet Bus
Message-ID: <1990Oct3.013509.1470@news.iastate.edu>
Date: 3 Oct 90 01:35:09 GMT
References: <211@bilpin.UUCP> <1990Sep28.020717.22610@dhw68k.cts.com> <1990Oct1.200613.635@tera.com>
Sender: usenet@news.iastate.edu (USENET News Poster)
Reply-To: carter@iastate.edu (Carter Michael Brannon)
Organization: Iowa State University
Lines: 55

>Shared memory systems are hardly technological dinosaurs.   The fact
>that bus contention becomes unmanageable in large systems just means
>that busses aren't a good idea.   There are a variety of other ways
>to implement shared memory.   Shared memory,  in fact,  can be
>demonstrated to be the only cost effective solution.

     Well Doc, that last statement was bound to draw a little bit of fire,
so here's the first dose! :-)  I'd like a qualification of the statement
that "Shared memory ... only cost effective solution."  While I must agree
with you that shared-memory systems are far from dinosaurs, I must disagree
categorically that they are the only cost-effective way to build a successful
massively-parallel computer.

     Here at the Scalable Computation Facility at Ames Lab, we're putting
together a little collection of machines.  One of the purposes of this
facility is to show that parallel computers of all flavors can be effective
(in MFLOPS and $/MFLOPS) toward most problems.  I qualify that statement with
"most" solely in deference to the relatively small number of ill-suited tasks
for certain groups of parallel architectures.  (Ray-tracing, for example, is
difficult on a SIMD machine, although it HAS been done with some success)
As a matter of fact, we see that quite a lot of problems don't care AT ALL
upon what architecture they are implemented (by any metric)!  I will grant
you that most of our applications are scientific applications -- lots of
number-crunching and not much I/O, but parallel disks go a long way toward
narrowing the gap with conventional mainframe I/O systems.

     The main thrust of this response is to provide a counterexample to the
non-cost effectiveness of massively-parallel distributed-memory computing.
We have an 8192 node MasPar MP-1 (SIMD, 4-bit PE's, 16KB mem/PE, toroidal
mesh interconnect) here upon which we routinely see in excess of 300 MFLOPS
** ON REAL PROBLEMS **!  If you subscribe to the $/MFLOPS school, that comes
out to roughly $720/MFLOPS!  Compare that with over $10,000/MFLOPS for your
favority Cray Y-MP.  The MasPar is certainly not the only example, just the
best one I can think of.  In fact, I can't think of ANY distributed-memory
parallel machines currently available that don't come in well under the
figures for the Cray Clique in terms of cost-effectiveness.

     I am much impressed with the products offered by Sequent and Encore
(Unfortunately, I am not familiar with TerraData's products.)
in terms of supporting hoards of users running vi, cc, and all that rot,
but neither can you deny that you PAY for that performance.  If your
definition of cost-effectiveness is something like $/user, then I will admit
that shared-memory mainframes have a considerable advantage over their
distributed-memory breatheren.  If, on the other hand, you're referring to
supercomputers, then I must champion the distributed-memory machine and state
that is just isn't the case.

Michael B. Carter             |  Violence is the last refuge of the
carter@iastate.edu            |  incompetent. -- Salvor Hardin
Scalable Computation Facility |
Ames Laboratory,              |
Ames Iowa                     |

--
Michael B. Carter