Thread View: comp.sys.transputer
1 messages
1 total messages
Started by On behalf of...
Tue, 11 Aug 1992 22:28
T9000
Author: On behalf of...
Date: Tue, 11 Aug 1992 22:28
Date: Tue, 11 Aug 1992 22:28
437 lines
18727 bytes
18727 bytes
PROF WELCH SENT ME THiS SO i THOUGHT i'D SHARE IT WITH YER ALL:
Dear All,
Please find enclosed a call for participants in two consecutive one-day
workshops in parallel computing. One is on the new INMOS T9000 transputer
and what it means for system builders (software & hardware). The other is
to initiate some real effort in re-targetting occam on to non-transputer
platforms.
I would be very grateful if you would circulate this information to all
interested parties in your departments.
Many thanks,
Peter Welch.
Professor of Parallel Computing, University of Kent.
(23rd. June, 1992)
(cut here)
-------------------------------------------------------------------------
T9000 SYSTEMS WORKSHOP
14-15 September 1992, University of Kent at Canterbury
AIMS
This one-day workshop is for all those who are planning (or, maybe, just
considering) to use T9000-based systems. It will especially examine
those features of the T9000 architecture that extend its functionality
beyond that provided by existing transputers (as opposed to those
features that merely make it go faster) and consider the impact of those
features on the way we design systems. In this context, `systems' mean
anything from a single processor embedded controller with no external
memory to numerical software for a general purpose 10,000 node super-
computer.
The claims of the T9000 architects are that it is not only a (10 times)
faster transputer than the existing generation, but also a (10 times)
`better' one - whose properties will greatly simplify systems' design and
programming. These benefits will enable very high efficiencies to be
easily obtained in multi-processor practice that will more than
compensate for any perceived shortfall in `peak' performance (that may be
apparent from paper comparisons against other chips).
This workshop will enable participants to hear the arguments behind the
T9000 design, to discuss and challenge those decisions and to develop an
early expertise in the engineering opportunities associated with the
device.
METHODS
Delegates will be presented with a detailed technical survey of the main
features of the T9000 transputer and C104 router. For the T9000, these
include the ALU/FPU architectures, super-scalar pipeline, hardware
instruction grouper, multi-level cache hierarchy, DS-links, virtual
channel processor, semaphore/shared-channel hardware, hardware scheduler,
external event pins, control links, external memory interface, central
cross-bar switch and error-trapping hardware. For the C104, they include
the 32-way parallel cross-bar, interval labelling, worm-hole routing,
random and universal routing, header deletion and network hierarchies.
As each feature is discussed, its practical implications for end-user
systems' design and programming will be emphasised. Questions to be
raised include :-
o the compute/communicate ratio for the T9000 seems close to
that for the T800. Does this imply that well-balanced T800
systems will remain well-balanced on T9000 hardware and
maintain similar performance-scaling properties?
o does the hardware instruction grouper imply that good
compilers need not be expensive and that programmers will
not have to craft careful assembler inserts to sustain high
efficiency in the processor pipeline?
o how much arithmetic per store access needs to be done to
enable the cache mechanisms to keep the processor running at
full speed? How does the cache hardware affect context-
switch times and interrupt response latency? How do we
decide whether to use full, half or zero on-chip cache?
o what is the real user-data bandwidth over DS-links? How is
this affected by the size of user-packets, the number of
user-channels sharing the link and C104 routers?
o how do we access the error-trapping and recovery mechanisms
(e.g. in the case of failure of a DS-connected device during
message transmission)?
o the T9000 virtual channel processor and the C104 router
promise a complete divorce of user-application design from
physical network topology. Is this true? If so, are there
any circumstances when it would still be better to connect
T9000s directly (e.g. for some embedded systems)?
o what simulations have been performed on large-scale
T9000/C104 networks and what conclusions can be drawn about
scalability and general-purpose parallel computing?
o theoretical studies have shown that efficient use of
general-purpose parallel architectures requires a degree of
parallelism in the software well exceeding that available in
the hardware. What implications does this hold for T9000-
based systems and how do they compare with other parallel
architectures on this issue?
o should access to the abstract higher-level parallel
processing model provided by the various T9000/C104 features
be gained through a programming language (e.g. occam3), an
operating system (e.g. Chorus, Helios, ...) or some mixture?
Participants are encouraged to bring along their own questions - the
above list is intended only as a `starter' pack. In addition to formal
presentations, the workshop will divide into smaller discussion groups
whose results will be presented back to the final plenary session. The
formal presentations will be led by senior design engineers from INMOS.
DATES, TIMES AND COSTS
Start: 2.00 p.m., Monday, 14th September, 1992.
Finish: 1.00 p.m., Tuesday, 15th September, 1992.
Cost: 70 pounds (includes lunch on Monday and Tuesday, dinner on
Monday evening, bed and breakfast, breaks and copies of lead
presenters' viewfoils).
ASSOCIATED OCCAM RE-TARGETTING WORKSHOP
The T9000 Systems Workshop is organised by the World occam and Transputer
User Group (WoTUG), in association with the (UK) ISC Initiative on
Software Environments for Parallel Computers at the University of Kent at
Canterbury and the COMETT technology transfer programme of the EC.
It is the first of two one-day workshops being held sequentially. The
other workshop is entitled ``Occam Re-Targetting Workshop'' and
immediately follows this one, starting at 2.00 p.m. on Tuesday, 15th
September, 1992. Further information on both workshops may be obtained
from :-
Professor P.H.Welch
WoTUG Chair Tel: +44-227-764000 ext. 7695
Computing Laboratory FAX: +44-227-762811
University of Kent email: phw@ukc.ac.uk
Canterbury, Kent, CT2 7NF
(cut here)
-------------------------------------------------------------------------
R E G I S T R A T I O N F O R M
---------------------------------
T9000 SYSTEMS WORKSHOP (14-15 September, 1992)
OCCAM RE-TARGETTING WORKSHOP (15-16 September, 1992)
University of Kent at Canterbury
Please tick against items required :-
|-----------------------------------------------------------|
|T9000 Systems Workshop 50 pounds [ ] |
|(including meals and breaks) |
|Bed and breakfast (14th September) 20 pounds [ ] |
|-----------------------------------------------------------|
|TOTAL FEE ENCLOSED pounds
|-----------------------------------------------------------|
|-----------------------------------------------------------|
|Occam Re-Targetting Workshop |
| |
|I enclose a brief description of my [ ] |
|motivation, expertise and possible |
|resources. Please send me an invitation |
|and special registration form. |
| |
|Costs will be the same as for the T9000 Systems Workshop. |
|-----------------------------------------------------------|
Booking: For either workshop, please complete and return this form.
If you are applying for the T9000 Workshop, please also
include:
either (a) a cheque (payable to UNIKENT) for items ticked
in the above box;
or (b) an official order form from your organisation.
Return to:
Judith Farmer
Computing Laboratory
University of Kent
CANTERBURY
Kent, CT2 7NF
Name:..................................................................
Position:..............................................................
Company:...............................................................
Address:...............................................................
.......................................................................
..................................... Post code: .....................
Daytime Tel. No:.......................................................
(cut here)
-------------------------------------------------------------------------
OCCAM RE-TARGETTING WORKSHOP
15-16 September 1992, University of Kent at Canterbury
AIMS
This one-day workshop is for people who are seriously considering re-
targetting the INMOS occam compiler to non-transputer platforms. This
includes those who intend to produce a compiler for purposes of education
in parallel computing and those who are considering a commercial occam
product. Educational platforms include PCs and workstations.
The workshop will be directed by key INMOS personnel to provide the
technical information and insight that will enable participants to re-
target the existing INMOS occam compiler to the platform of their choice,
to estimate the size of this task and to estimate the performance of the
resulting system.
Attendance to the workshop will be by invitation only. If you would like
to attend, please contact Peter Welch (address below) giving a brief
indication of your motivation, your expertise and the amount of resource
you may be able to devote to such a project.
MOTIVATION
Occam is fast becoming the `de-facto' standard for the teaching of
parallel processing concepts and practice. The high-level visibility it
provides for the structure and dynamics of message-passing systems,
together with the security and stability it enforces through strongly
policed design rules, makes it the ideal language for building
confidence, fluency and insight into parallel computing. These
properties also make it the language of choice for a growing spectrum of
application areas: from high-performance embedded systems (especially
those for which safety-critical issues are dominant) through to numerical
super-computing.
These developments are now being constrained by the `transputer-only'
label attached to the language. Most higher educational institutions do
not provide enough transputer-based hardware to support the `mass'
teaching of parallel computing - for instance, as a compulsory subject
for first-year undergraduates (where class enrolments can be well over
100 students). A successful re-targetting of the occam compiler to a
workstation or PC environment would remove this constraint and allow the
whole computing community to experiment seriously with concurrency.
WHAT IS INVOLVED
Compiler Front-End:
For re-targetting the occam compiler to a non-transputer architecture, it
is strongly recommended that the front-end of the existing compiler be
re-used without alteration. This will not only save much time and ensure
that full semantic checking (e.g. of alias, usage and range rules) is
correctly and efficiently applied, but will enable re-targetted compilers
to keep track of future developments in the primary transputer compiler.
The output from the INMOS front-end is an abstract syntax tree.
Attendees of the workshop will be given a brief introduction to the INMOS
compiler, together with guidelines about how best to re-use this
implementation.
Realising a Communicating Process Model:
Most of the work in re-targetting the compiler will be involved in
deciding how to exploit a particular platform (i.e. hardware and
operating system combination) in order to realise a communicating process
model. In general, this will be done through a mixture of code
generation and run-time kernel implementation. Attention must be given
to achieving acceptable performance for context switches, process start-
up and message latencies and communication bandwidth (e.g. for carrying
synchronised occam channels across LANs).
It is hoped that co-ordination between occam re-targetting efforts will
be able to produce a standard interface to the communicating process
run-time system.
INMOS OCCAM POLICY
Occam is a registered trade-mark of the INMOS group of companies.
However, INMOS are actively encouraging the re-targetting of the language
to non-transputer architectures. Source code and documentation for the
INMOS occam compiler are available as a product from their distributors
(IMS D5202A-S, about 125 pounds, and comes on a QIC-24 cartridge tape).
It is, of course, essential that control be maintained on the language
(to prevent a proliferation of dialects or extensions that break the
principles of occam). Licensing arrangements to ensure this will be
discussed at the Workshop. In the longer term, INMOS is seeking to pass
this control to the relevant international standards authority.
DATES, TIMES AND COSTS
Start: 2.00 p.m., Tuesday, 15th September, 1992.
Finish: 1.00 p.m., Wednesday, 16th September, 1992.
Cost: 70 pounds (includes lunch on Tuesday and Wednesday, dinner on
Tuesday evening, bed and breakfast, breaks and copies of lead
presenters' viewfoils).
Attendance: By invitation only! Please see the AIMS section above.
ASSOCIATED T9000 WORKSHOP
The Occam Re-Targetting Workshop is organised by the World occam and
Transputer User Group (WoTUG), in association with the (UK) ISC
Initiative on Software Environments for Parallel Computers at the
University of Kent at Canterbury and the COMETT technology transfer
programme of the EC.
It is the second of two one-day workshops being held sequentially. The
other workshop is entitled ``T9000 Systems Workshop'' and immediately
precedes this one, starting at 2.00 p.m. on Monday, 14th September, 1992.
Further information on both workshops may be obtained from :-
Professor P.H.Welch
WoTUG Chair Tel: +44-227-764000 ext. 7695
Computing Laboratory FAX: +44-227-762811
University of Kent email: phw@ukc.ac.uk
Canterbury, Kent, CT2 7NF
(cut here)
-------------------------------------------------------------------------
R E G I S T R A T I O N F O R M
---------------------------------
T9000 SYSTEMS WORKSHOP (14-15 September, 1992)
OCCAM RE-TARGETTING WORKSHOP (15-16 September, 1992)
University of Kent at Canterbury
Please tick against items required :-
|-----------------------------------------------------------|
|T9000 Systems Workshop 50 pounds [ ] |
|(including meals and breaks) |
|Bed and breakfast (14th September) 20 pounds [ ] |
|-----------------------------------------------------------|
|TOTAL FEE ENCLOSED pounds
|-----------------------------------------------------------|
|-----------------------------------------------------------|
|Occam Re-Targetting Workshop |
| |
|I enclose a brief description of my [ ] |
|motivation, expertise and possible |
|resources. Please send me an invitation |
|and special registration form. |
| |
|Costs will be the same as for the T9000 Systems Workshop. |
|-----------------------------------------------------------|
Booking: For either workshop, please complete and return this form.
If you are applying for the T9000 Workshop, please also
include:
either (a) a cheque (payable to UNIKENT) for items ticked
in the above box;
or (b) an official order form from your organisation.
Return to:
Judith Farmer
Computing Laboratory
University of Kent
CANTERBURY
Kent, CT2 7NF
Name:..................................................................
Position:..............................................................
Company:...............................................................
Address:...............................................................
.......................................................................
..................................... Post code: .....................
Daytime Tel. No:.......................................................
Thread Navigation
This is a paginated view of messages in the thread with full content displayed inline.
Messages are displayed in chronological order, with the original post highlighted in green.
Use pagination controls to navigate through all messages in large threads.
Back to All Threads