jww@philabs.Philips.Com (James Wendorf) (03/15/88)
I only received 3 responses to my earlier query about 68020/30 based multiprocessors. I have included them below for your information, along with my original request. Any further pointers would also be appreciated. My thanks to those who responded. -Jim Wendorf ------------------------------------------------------------------------------ Subject: 68020/30 based multiprocessors? From: jww@philabs.UUCP (James Wendorf) We are seeking information on commercially available multiprocessors, based on the 68020/30. In particular, we are interested in machines with coherent, uniformly accessible, shared memory. We are looking into Computer X, but apparently that system won't be available until the end of this year. Has anyone constructed a system (with caches) from commercially available boards? I'll summarize any interesting replies for the net. ------------------------------------------------------------------------------ From: harvard!ll-xn!lll-crg!lll-tis!viusys!rwb@linus (Rick) Organization: Unisys D.A. MINIS Branch, McLean, VA Date: 26 Feb 88 10:07:43 EST (Fri) The Unisys 5000/85/90/95 machines are MC68020-based machines, with symmetrical multi-processing, (not parallel). Memory is global, with no private RAM, other than cache, (not Megaframe mentality). Briefly, the specs are: 5000/85 - Tower sized, up to two 25MHz MC68020 CPU's, each with 12.5MHz MC68881, IPL proms, 64 KB cache, MMU. 5000/90 - Larger Cabinet, up to 4 12.5 MHz MC68020 CPU's, each w/ 12.5 MHz MC68881, IPL proms, 8 Kb cache, MMU. 5000/95 - Same cabinet as 5000/90, but with up to two of the 25 MHz flavored CPU boards. The above systems run V.2/3, the 12.5 boards support 16 MB main, the 25 supports 64 MB. There is a master CPU designated in these systems, which is switch selectable. Slave CPU's can access run queues, and other system structures, and can execute most system calls (~95%) so it's not the bottleneck you might think. The MC68030 architecture is such that it's almost certain that there will never be a multi-processor system based on it. Rick Butland (rwb@viusys) ------------------------------------------------------------------------------ From: sun!portal!cup.portal.com!DGD@nynex1 Date: Sat Feb 27 18:36:36 1988 Honeywell BULL Italy makes such a machine. I think it's called the XPS-100 Model 40. Or is that XPS 1000? ------------------------------------------------------------------------------ From: amdcad!ironics@gatech (Paul Davis) Organization: Ironics Inc. (c/o AMD) Date: Tue, 1 Mar 88 13:43:39 PST You can build a 20 processor box with standard VME hardware in one chassis. Using some of the various approaches to inter-VMEcrate connect you can get up to number like 100+ processors in one rack. VME (using A32 addressing) gives 4GB of shared memory space given the right hardware choices. You can also get BIG (up to 1GB) memory array boxes that can be attached. Also, JPL has some big arrays of 020's built -- I don't know much more than they are hypercube based and that there is some kind of commercial arrangement to sell it if it works. Paul Davis, Ironics Inc., Ithaca, NY Reply: ironics@amdcad.AMD.COM Voice: (607) 277-4060 ------------------------------------------------------------------------------
jww@philabs.Philips.Com (James Wendorf) (04/19/88)
I received several more responses to my earlier query about 68020/30 based
multiprocessors. All the responses I received have been included below for
your information, along with my original request. Many thanks to all those
who responded.
One system we are seriously considering, but not mentioned in any of the
responses, is the BBN Butterfly. It uses 68020/4MB processing nodes connected
by a multistage switching network, and can have up to 256 nodes. For more
information, contact:
BBN Advanced Computers Inc.
10 Fawcett Street
Cambridge, MA 02238
(617) 873-6000
-Jim Wendorf
------------------------------------------------------------------------------
Subject: 68020/30 based multiprocessors?
From: jww@philabs.UUCP (James Wendorf)
We are seeking information on commercially available multiprocessors, based on
the 68020/30. In particular, we are interested in machines with coherent,
uniformly accessible, shared memory. We are looking into Computer X, but
apparently that system won't be available until the end of this year. Has
anyone constructed a system (with caches) from commercially available boards?
I'll summarize any interesting replies for the net.
------------------------------------------------------------------------------
From: harvard!ll-xn!lll-crg!lll-tis!viusys!rwb@linus (Rick)
Organization: Unisys D.A. MINIS Branch, McLean, VA
Date: 26 Feb 88 10:07:43 EST (Fri)
The Unisys 5000/85/90/95 machines are MC68020-based machines, with symmetrical
multi-processing, (not parallel). Memory is global, with no private RAM, other
than cache, (not Megaframe mentality). Briefly, the specs are:
5000/85 - Tower sized, up to two 25MHz MC68020 CPU's, each with 12.5MHz
MC68881, IPL proms, 64 KB cache, MMU.
5000/90 - Larger Cabinet, up to 4 12.5 MHz MC68020 CPU's, each w/ 12.5 MHz
MC68881, IPL proms, 8 Kb cache, MMU.
5000/95 - Same cabinet as 5000/90, but with up to two of the 25 MHz flavored
CPU boards.
The above systems run V.2/3, the 12.5 boards support 16 MB main, the 25
supports 64 MB. There is a master CPU designated in these systems, which is
switch selectable. Slave CPU's can access run queues, and other system
structures, and can execute most system calls (~95%) so it's not the
bottleneck you might think.
The MC68030 architecture is such that it's almost certain that there will
never be a multi-processor system based on it.
Rick Butland (rwb@viusys)
------------------------------------------------------------------------------
From: sun!portal!cup.portal.com!DGD@nynex1
Date: Sat Feb 27 18:36:36 1988
Honeywell BULL Italy makes such a machine. I think it's called
the XPS-100 Model 40. Or is that XPS 1000?
------------------------------------------------------------------------------
From: amdcad!ironics@gatech (Paul Davis)
Organization: Ironics Inc. (c/o AMD)
Date: Tue, 1 Mar 88 13:43:39 PST
You can build a 20 processor box with standard VME hardware in one
chassis. Using some of the various approaches to inter-VMEcrate
connect you can get up to number like 100+ processors in one rack.
VME (using A32 addressing) gives 4GB of shared memory space given
the right hardware choices. You can also get BIG (up to 1GB) memory
array boxes that can be attached. Also, JPL has some big arrays of
020's built -- I don't know much more than they are hypercube based
and that there is some kind of commercial arrangement to sell it if it
works.
Paul Davis, Ironics Inc., Ithaca, NY
Reply: ironics@amdcad.AMD.COM
Voice: (607) 277-4060
------------------------------------------------------------------------------
From: mcvax!cernvax!hve@uunet (H.v.Eicken)
Organization: CERN European Laboratory for Particle Physics, CH-1211 Geneva, Switzerland
Date: Thu, 17 Mar 88 09:27:39 +0100
You might want to contact:
Creative Electronic Systems SA
70, Route du Pont-Butin
PO Box 107
CH 1213 Petit-Lancy 1
Geneva, Switzerland
Tel. (022) 925745
Tlx: 421 320 CES-CH
Fax: (022) 925 748
Their VME based FIC 8230 board, currently with MC 68020, might be interesting
for you. They have multiprocessor systems up and running with PSOS as
operating system.
Horst von Eicken, DD-Division, CERN, Geneva 23
------------------------------------------------------------------------------
From: rbl%nitrex@mandrill.CES.CWRU.Edu
Date: Fri, 18 Mar 88 14:13:25 EST
The problem with paralleling 680X0's is the cache coordination issue. I'm
only aware of a few RISC processors (Berkeley's SPUR and perhaps the SUN
SPARC) that handle it well. I recall ENCORE doing it with National's
32XXX chip, too.
I was at Berkeley last week and they are submitting their SPUR design to
silicon via "MOSES" -- which is some sort of silcon foundry.
Counterpoint Systems was supposed to produce a multiprocessor product
for AT&T about a year ago, but that deal fell thru. I don't know what
happened to it or if it was 680X0-based.
Rumor at Berkeley was that Motorola's new RISC chip will run at 80 MHZ,
produce close to 80 MIPS and will have a cache coordination capability
with off-CPU chips. An 80 MHz CMOS version of the 680X0 is also
rumored.
Rob Lake
BP America R&D
mandrill.CWRU.EDU!nitrex!rbl
------------------------------------------------------------------------------
From: Dave Forsey <watmath!watcgl!drforsey@uunet>
Organization: U. of Waterloo, Ontario
Date: Fri, 18 Mar 88 11:35:31 EST
Must have missed the original posting. Dy4 systems out of Ottawa canada
makes multiprocessor VME systems. Up to 16 68020 + fpu (16 slots, but you
do need I/O boards, so you don't put in 16) with hardware to assist for message
passing. Runs the Harmony operating system - message-based (synchronous
send/receive/reply) real-time. Not unix. Compile on host and download
(can be done over ethernet). We run several of these systems in the lab here.
Vertigo computer graphics of Vancouver also make a multiprocessor system
using 68020's (two per board) - they were bought out by Cubicomp.
Two other companies (forget the names) also have Harmony-based systems.
One is associated with Carlton University in Ottawa, the other is in
Toronto.
The Harmony O/S was developed by Morven Gentleman at the National Research
Council in Ottawa, Canada.
Dave Forsey
Computer Graphics Laboratory
University of Waterloo.
------------------------------------------------------------------------------
From: mcvax!nlgvax!tom@uunet (Tom v. Peer)
Organization: Philips Research Geldrop
Date: Tue, 22 Mar 88 16:39:51 +0100
Philips I&E should be coming out with their 68020 VME/VMX boards by June
this year.
You can run their DRM system on it.
DRM stands for Distributed Real-time Monitor : an operating system with
"Real-time" capabilities which supports multi processing.
Each board has its own private memory (2Mb or 4Mb, I don't know) which is
accessible from the bus for other processors. It's also
possible to have shared memory but in DRM communication between processors/
processes is done using message-passing.
Tom van Peer
Nederlands Philips Bedrijven
Dependance Nat. Lab. Building XR
Willem Alexanderlaan 7B
5664 AN Geldrop - The Netherlands
Tel: +31-40-892334
E-mail: mcvax!philmds!nlgvax!tom
peer@vangogh.prl.philips.nl (may vanish soon)
------------------------------------------------------------------------------
From: kansas!curtg@masscomp
Date: 24-Mar-1988 15:21-EST (Thursday)
MASSCOMP has been building multiprocessor 680X0 Unix systems for 4
years now. We currently have 680X0 systems from $15k-$300k based
around 1-8 16MHz 68020 or 1-4 25MHz 68030. Our primary target market
is realtime and data-acqusition but we also sell into the general
Unix workstation market. We have a wide range of peripherals and
graphics options for the Multibus or VMEbus. For more information
contact:
MASSCOMP
One Technology Way
Westford, MA 01886
(617)-692-6200
If you have any specific questions about our systems I would be glad
to try to answer them via email.
curt gridley
principal engineer
MASSCOMP
...!{decvax,ihnp4,seismo}!masscomp!curtg
------------------------------------------------------------------------------
<<<Received via physical mail.>>>
From: wien@nrcctis.UUCP
We have been using single board computers form DY-4 Systems Inc for our
multiprocessors in the development of our Harmony operating system and for
robotics research.
Packaged systems (as well as Harmony) are available commercially from P-Can
Research Inc:
P-Can Research Inc.
80 Galaxy Blvd., Unit 7
Rexdale, Ontario
Canada M9W 4Y8
(416) 674-6600
We have been using Macs for software cross-development. They are using Suns.
Marceli Wein
Division of Electrical Engineering
National Research Council of Canada
Ottawa, Ont Canada, K1A 0R8
------------------------------------------------------------------------------