🚀 go-pugleaf

		    WoTUG T9000 Systems Workshop

		  University of Kent at Canterbury

		       14-15th September 1992


  This is a quick summary of the proceedings of the T9000 systems
workshop given at the University of Kent on the 14th and 15th of
September 1992 presented by the World occam and Transputer User Group
(WoTUG) in association with COMETT and the Software Environments for
Parallel Computing (SEPC) Initiative.

  Roger Shepherd (RMS) and Peter Thompson (PWT) from INMOS Ltd,
Bristol, UK were the invited speakers.

  The first day began with an overview of what the T9000 offered,
presented by RMS. Firstly he described virtual channels, routing and
the performance of the channels.  He said that at 100Mbit/s, the
links could do 9.55Mbytes/s unidirectional and 2x 8.74Mbytes/s
bidirectional.  He then went on to discuss the memory system, how the
cache banks were connected up to the CPU, Scheduler and VCP, the
processor pipeline including instruction grouper, the process model
and the configuration which can be done via software.  Finally he
described the control links and their use.

  The second talk was given by PWT in which he gave an introduction
to the communications architecture of the T9000.  Comparisons were
made between T805s and T9000s showing how the hardware
implementations of variable-length communications, multiplexing and
virtual channels cause large performance increases.  He also
described how message routing could be used and how the T9000/C104
pair make decoupling software and hardware much easier.

  The next talk was given by RMS and covered the pipeline system of
the T9000.  How the instruction grouper worked was covered in detail
with examples.  The result is several instructions can be done per
cycle without the need for very clever optimizing compilers.  However
he did point out that a good T9000 compiler still needs to know about
the grouping to achieve the best results.

  In the evening PWT described the memory system, the cache behaviour
and device access.

  On the second day, the first talk was given by PWT and described
the communications protocol in detail.  The protocol is layered,
including physical, token, packet and message layers.  The token and
packet types and operation were described in great detail and the
workings of the virtual links presented.  The links have a low level
flow control system using tokens and parity checks.  A presentation
of the calculations for visible bandwidth, taking into account the
protocols was given (the results were given above).  Graphs of the
data throughput versus messages size showed that for reasonable size
messages, the data throughput was good.  The message routing system
(Wormhole Routing) which uses interval labelling, addressing and
header deletion was described.

  The final talk, given by PWT, discussed errors in systems:
detection of errors and taking action after them.  He went on to
describe how the T9000 provides a P-process or "protected" process
that runs under the supervision of another process which gets called
when a privileged instruction is called or it tries to change global
state.  As an example he described an ATM switch made from a C104
network with T9000s interfacing to the outside world.

  The workshop concluded with questions from the floor about the
T9000.  A T9000 alpha silicon (with the VCP absent) was passed around
the room, courtesy of INMOS Japan.  The VCP is on the beta silicon
and a photograph of the processor die of it was shown.

  Several questions were asked at the end to the presenters and also
to Steve Doyle of INMOS marketing about the T9000.  Steve said that
in the short term there will be no support for mixed T2/4/8 and T9000
networks and the network description language used to configure
processor networks will be upgraded to handle virtual links.


Transcribed by David Beckett (djb1@ukc.ac.uk)
           and Vedat Demiralp (svd@ukc.ac.uk)

There will probably be more information available soon,  watch this
space!

October 14th 1992