[comp.arch] SUMMARY: multiprocessor systems without shared-memory

bwong@ihwpt.ATT.COM (bruce wong) (05/27/88)

First, a repeat of part of my second posting:

Subject: General purpose commercial multiprocessor systems without shared-
memory

By general purpose multiprocessor system I mean that the machine looks like
a uniprocessor machine to the user at the shell level and at the application
programming level.  The user is not concerned at all about processor allocation
and application partitioning.  This probably means that parallelism occurs at
the process level.

These are the commercial systems which fit the requirements:

	AT&T 3B4000
	ELXSI
	NCR TOWER 32/800
	Various Transputer machines

As you will notice below, I don't have much information on the ELXSI.
Is it still in business ?

I also got mail about experimental systems.  I have included some of those
below.
#-----------------------------------------------------------------------------#
I assume that you, by now, have heard of the AT&T 3B4000 (though our
Marketing people are doing a good job of keeping it a secret!).

It is a multiprocessor (up to 14 adjunct processors in addition to the
Master Processor) which looks from the outside (at the shell and
application level) like a single UNIX machine.  Each processor has the
full set of SVR3.1.1 shared memory primitives, but a shared memory
segment cannot (yet) be shared across processors.  Hence, there is
local shared memory (between processes running on the same processor)
and no global shared memory (between processes running on different
processors).

Incidently, this restriction is consistent with the SVID, which
identifies shared memory as a machine-dependent feature.  In case you
would like a closer look, the machine ihaxa at the IHCC is a 3B4000.

				Rich Strebendt
				[iwsl6|ihlpe|ihaxa]!res
				[cuuxf|cuuxg]!iw1res!res
#-----------------------------------------------------------------------------#
I believe that ELXSI use messages, and don't have shared memory.
--
	Don		lindsay@k.gp.cs.cmu.edu    CMU Computer Science
#-----------------------------------------------------------------------------#
The NCR TOWER 32/800 implements a loosely-coupled multiprocessor system
with separate Application, File, Terminal, Communications and Ethernet
processors that are connected by Multibus II.  Each processor has its
own private memory, although cross-processor memory accesses are also
allowed.  Further, the system may be configured with multiple Application
processors that implement parallel processing in the method you describe --
that is, on the process level.  No special application coding is needed 
to exploit the feature, and it is transparent to the shell and to users.
Currently runs AT&T V.2.2.

Steve Wingard                        ...!ucbvax!sdcsvax!ncr-sd!ncrcae!wingard
NCR Corp., E&M-Columbia              ...!decvax!mcnc!ncsu!ncrcae!wingard
TOWER Muliprocessing Systems             wingard@ncrcae.NCR.COM
#-----------------------------------------------------------------------------#
We have several customers who build, and market, non-sharded memory,
general purpose multiprocessor computers, built with the (INMOS)
transputer as the processing and communication element.

Some such vendors are Meiko, THORN-EMI, Parsytech (not sure about the spelling),Topologix. I've left out Floating Point Systems because I feel that the
T-series is a rather specialised computer (it consists of a number of
interconnected vector processors).

As one example of an installation of a non-shared memory
multiprocessor computer, Edinburgh University (Scotland) have a large
Meiko machine, with about 200 processors currently, which supports
multi-users, networked access etc etc. (The processors
are mainly IMS T800s, which mean that the whole system has a 
sustainable performance of somewhere in the 300-400 MFlop region!). The
Edinburgh machine has been used for a wide range of application areas.
There are plans (subject to raising the necessary funds) to expand
this system to around a 1000 processors I believe.

I hope this is of use, and I ought to appologise to any of our customers
I have insulted by omitting them from this list.

Roger Shepherd
INMOS Ltd
1000 Aztec West
Bristol
BS12 4SQ, Great Britain   ....!mcvax!euroies!shepherd
#-----------------------------------------------------------------------------#
Ah you mean automatic partitioning, etc.  Well, I don't know of any turn
key parallel systems, but if users are willing there are a host of near
raw machines, mostly shared memory which have some neat features:
Sequent, Encore, Flex, etc.  But this does not constitute an
endorsement.

In the organization I just left, we have been approached by DEC on their M31
project (as were many other people) with parallel VMS (See COMPCON'88), two
different IBM projects, and numerous startups.  Remember a processor in
hand is worth 1,000 in the simulator.

Another gross generalization from

--eugene miya, NASA Ames Research Center, eugene@aurora.arc.nasa.gov
#-----------------------------------------------------------------------------#
From: ihnp4!thumper!hammond (Rich A. Hammond)

I know this isn't commercial, yet, but here at Bell Communications Research
there are two multi-processor versions of UNIX using either Ethernet
or a Ring to connect machines that appear to the user as a single
machine without processor boundaries.  Both support process migration
between processors and handle signals, ... properly across mutliple
processors.  At least one version could become commercial strength with
very little effort.

Rich Hammond
#-----------------------------------------------------------------------------#
-- 
	Bruce F. Wong	1A-111		312-416-5111
	ATT Bell Laboratories		ihnp4!ihwpt!bwong
	200 Park Plaza
	Naperville, Ill 60566-7050