[comp.edu] Summary of Multiprocessor Hunt

buz@cs.nps.navy.mil (09/08/89)

The following are a brief synopsis of the machines that I received
comments about.  The comments ranged from detailed testamonials
to breif mentions.  The relevant context from my original posting
is included immediately below.

==================== excerpt from original posting ====================
...I would like to augment this
very reasonable computing base with a parallel processor to:

(1) handle general CPU and disk intensive computing needs (i.e.,
    provide an efficiently managed pool of CPU cycle and disk resources)
(2) allow us to credibly teach parallel algorithms, architecture and
    parallel/distributed O.S. courses with realistic projects
(3) allow research experimentation with parallel algorithms,
    simulations, operating systems, etc. (both running on the native
    O.S. and via simulations of "virtual" parallel machines that we can
    build on top of the native O.S.).
(4) handle critical departmental computing needs in a reliable
    fashion (hopefully, without comflicting with (3) too much).

As a minimum the machine needs to:

(1) provide excellent price/performance
(2) have a good reliability and service reputation
(3) run an "almost perfectly real" version of 4.3BSD Unix (i.e. it must
    be an easy port target -- Mach would also be an interesting option)
(4) run tcp/ip, NFS, X11R3, tex/latex, lisp, ada, and other common academic
    software
=========================== end of excerpt ============================

Small scale multiprocessors (up to 8):
MIPS Co.	MIPS R3000 based multiprocessor server, good comments
		about compiler and NFS
Solbourne	Sparc based multiprocessor server, Sun binary compatible
Pyramid		CISC (not sure which) based multiprocessor, 4.3BSD
SGI		MIPS R2000 Sys V based
Ardent		MIPS based multiprocessor
Stellar		Proprietary multi-CPU design
Apollo		DN/DPS10000 68020 binary compatible (good recommendations
		on O.S. software, even)

The general consensus (with an exception or two) is that the
machines in this group reflect individually better (in price/performance
terms) CPU designs than those in the following group (i.e., many people
like the MIPS/Sparc/DN10000 CPU architectures)

Medium scale (8-32):
Sequent		80386 and 32032 versions, run Dynix (a reportedly good 4.3BSD
		sys V merge), mostly good comments, several very satified
		customers, some thought SW was poor, wide software base
Encore		good comments, Mach is available, wide academic/research
		software base
Alliant		68020 instr set, Can do loop parallelism, good number cruncher,
		probably not as good for supporting lots (100s) of users as
		Sequent or Encore, unsure of "general purpose" software base
Elxsi		Few comments, good number crunching capability (good
		optimizing Fortran compiler), unsure of "general purpose"
		software base (CPU number-wise this probably should
		have been in the preceeding group, but price-wise
		it fits better here)
Convex		Few comments, good number crunching capability (good
		optimizing Fortran compiler), unsure of "general purpose"
		software base

In this group the (apparent) consensus was to check out both Sequent
and Encore for large-scale (i.e. supporting many users) general purpose
computing needs, and to check out Alliant, Convex and Elxsi for fast number
crunching capabilities.

Unconventional architecture:
Multiflow Trace		VLIW machine, brief mention, easy to port target
			for common math packages, good speedups on fine
			grain parallel programs.  No comments on the
			suitability for general purpose computing.

Large scale multiprocessors (10s to 100s):
AT&T pixel machine	60 DSP style processors (didn't get many details)
BBN Butterfly		Good FP and/or Vector unit, new machine
			has much improved I/O and is M88000 based,
			older ones are 68020 (system can be grown
			to large/expensive size), Mach'ish operating
			system.

Parting Comments:
	We ended up using our money to upgrade existing equipment and to buy
	new software.  Thus, our purchase of new hardware will be delayed
	slightly.  Based upon the input that I got (some machines I
	didn't get many details on) the "good bet" awards go to SGI (for
	out-and-out MIPS/$ look at the 8 processor R3000 server) and BBN
	(for putting a LOT of general purpose processors in a reasonably
	priced, expandable, package).  These are the two (diverse) machines
	that I thought offered the best deal for the type of needs that we
	currently have.

Disclaimer:
	All of the usual disclaimers apply.  I make no guarantees
	for the thoroughness or completeness of information.  This
	summary reflects my opinions (based on net-land and vendor
	input) of the machines that I considered for our specific
	needs -- it does not reflect the opinions of my employer,
	God, Elvis (whereever he may be) or anybody else.

Thanks:	To all who supplied information.

Greg Buzzard (prof.) 			internet:  buz@cs.nps.navy.mil
CS Department (mail code 52Bu)
Naval Postgraduate School		phone:     408/646-2693
Monterey, CA  93943

eugene@eos.UUCP (Eugene Miya) (09/08/89)

Very interesting.  No one recommended or even mentioned the FLEX/32,
and I know of at least 3 sites (none net accessible).

Another gross generalization from

--eugene miya, NASA Ames Research Center, eugene@aurora.arc.nasa.gov
  resident cynic at the Rock of Ages Home for Retired Hackers:
  "You trust the `reply' command with all those different mailers out there?"
  "If my mail does not reach you, please accept my apology."
  {ncar,decwrl,hplabs,uunet}!ames!eugene
  				Live free or die.

dennism@menace.rtech.COM (Dennis Moore (x2435, 1080-276) INGRES/teamwork) (09/09/89)

Another un-mentioned multiprocessor:

A *great* price/performance machine is the system from Corollary (and SCO,
I think).  It's a symmetric multiprocessor running Xenix or Unix V.  I think
you can run any MS-DOS or 386 Unix software on it.

-- Dennis Moore, my own opinions etc etc etc