merlin@hqda-ai.UUCP (David S. Hayes) (02/27/88)
Well, my last query (about the Ultramax) managed to get me
interested in the Multimax. This appears to be a product that is
in current production, while the Ultramax is still an R&D item.
HAS ANYONE ACTUALLY USED THE MULTIMAX?
What experiences can you share? My interest is primarily in
it's ability to do parallel processing. We do not need to support
lots of Unix users. We would like to bring a great deal of
processor power to bear on a single program. Can the Multimax do
this? Can the compilers parallelize (is that a word?)
automatically, or do we have to tell them when to compile for
parallel execution? What else do we need to watch out for?
Thanks again for whatever information you have,
--
David S. Hayes, The Merlin of Avalon PhoneNet: (202) 694-6900
UUCP: *!uunet!cos!hqda-ai!merlin ARPA: ai01@hios-pent.arpabzs@bu-cs.BU.EDU (Barry Shein) (02/28/88)
From David S. Hayes
> HAS ANYONE ACTUALLY USED THE MULTIMAX?
We have 7 MultiMaxes at Boston University. I'd say someone here has
used them. We're currently using them in various environments, the
majority being some form of time-sharing (3 are used for course-work,
Computer Science, Engineering, Adult Ed, others are in VLSI design
labs, a school's administrative database, a departmental general
purpose system which eventually should be supporting a bunch of Suns.)
As far as parallelizing code, the current tools require coding
explicit parallelism via the use of library calls designed for this
(tasks or threads creation and manipulation, semaphores, spinlocks
etc.) I believe there is something in the works for automatic
parallelization, you should probably call them (617-460-0500), I'm
sure they'd be happy to chat about your specific applications. They do
have some tools within the Unix environment, parallel make, grep,
things like that, but I suspect that's not what you're asking about.
I've written and benchmarked a few applications using the parallelism
of the machine (a ray-tracer and a sort routine.) I found the speed-up
to be encouraging (benchmarking the process from 1..N CPUs.) I didn't
have to do anything terribly exotic in the code but they were easily
partitioned tasks (I'm no fool), just splitting up arrays and starting
sub-processes to each do part of the job. I suspect this is the most
attractive sort of problem to attack with hand-coding, usually only
requires dividing the array by the number of CPUs you intend to use
and assigning index ranges to each sub-process. Basically a for() loop
which starts the sub-processes on each segment. The sort needs a merge
phase at the end, but all standard kinds of things.
More importantly, you never begin to understand the paradigm until
you begin using it.
-Barry Shein, Boston Universitywtm@buengc.BU.EDU (W. Thomas Meier) (03/03/88)
I am a system manager on a few of the 7 Multimaxes here at B.U.. Most of
our systems are DPC CPU systems, but I would recommend the APC Processors
because of their greater speed.
We have had some problems with reliability with our systems, but Encore
has been very "energetic" in bug fixing from both the software and hardware
point of view. We have been getting 12 to 16 days of uptime. QA on
hardware is fair, but getting better.
The average configuration here at B.U. Running Umax 4.2 is
3 DPC cards (6 cpus)
1 EMC card (ethernet communications, etc.)
1 SCC card (system control card, can't do without)
1 LDC-01 (disk controller card)
2 CDC disks (about 400Meg each formatted)
and a gaggle of Annex boxes for serial line to ethernet communications;
the Multimax has no serial line ports except for the console and a diag
port.
If I can be presumptuous for a second, a recommended configuration
based on our usage (especially with Ada), might be:
3-5 APC cards (6 to 10 cpus)
1 EMC card
1 SCC card
1 LDC-02 (is 2 physical ldc-01s on the same card for more through-put
when adding to your first two disks)
2 NEC disks (about 600Meg each formatted)
_W. Thomas Meier -Boston University / Distributed Computing & Communications
"... I wonder if Encore gives the U.S. Government discounts"m4@UX.ACSS.UMN.EDU (Marshall M. Midden) (03/03/88)
Our machine is used mainly for student classroom work. From "an overview of many programming languages", to statistics and vlsi. We average about 25 people logged on at a time (8am - 1am), about 12 on Saturday and Sunday. For this type of usage, I recommend lots of everything! We have the following: 3 APC cards (6 processors), 64MB of memory, 7 CDC disks [415MB], 4 NCR disk controllers, 2 CDC shoe-shine tape drives (both masters), 1 Fugi 200IPS start/stop master tape drive (2nd arriving any day), 5 EMC cards: 1 for each disk controller; 1 for the Fugi, +1 any day; the 2 cdc tape drives are on two containing disk controllers; we use one ethernet for annexes solely, and another for the campus ethernet. There is a power distribution unit so things can be moved relatively easily in the machine room (except for the Fugi's). The two CDC tape drives will be returned someday, shortly after we get the 2nd Fugi. We will still need 2 EMC cards, but when the MSC is available, the other 3 (or 4 if the Fugi comes first) will be returned. We hope to go from 3MB disk storage to 6MB, and add an addition 2 APC cards (4 processors) for next fall. The release 3.1 of the software is very very stable for our environment. We have been up 26 days since the last power failure. There are minor software complaints (UMAX and annexes), but everything serious has been fixed, just a bunch of annoying bugs and wierdness.
jb@CS.BROWN.EDU (Jim Bloom) (03/04/88)
The computer science dept. at Brown Univ. has a 12 processor machine. The machine is being used primarily for research, but there is a class doing some work with parallel programming using it (ten students). We generally have only a few users at a time. The machine is generally reliable. The software has a few bugs (but I haven't seen any vendor yet that didn't have some). One research project is working on supporting parallel computation. We have developped a system for using multiple processors to perform a calculation. Our work has yield a package which provides a runtime system for parallel computation. This code is currently a library that could easily be used to provide the basis of a parallelizing compiler. As far as writing programs, the programmer must understand where the possibility of parallelism exists and how to make the best use of it. C is supported currently and Fortran is very close to completion. Also, this currently only works under the BSD based system and not the Sys V version. To answer your questions more directly: Yes, the multiple processors may be used for a single program. I do not know of any compilers available from Encore that parallelize the code. As such, you will need to do it yourself. Our code will be available from Encore in the near future. Jim Bloom Brown University
"Jim@UCBVAX.BERKELEY.EDU (03/07/88)
I'm not sure why I've started receiving mail about Encore Multimax, but I would guess that I've been added to your mailing list by mistake. Please remove my name. Thanks, Jim F.