Parsytec GmbH
Parsytec GmbH
Status:
Overview of Organisation:
Parsytec (PARallel SYstems TEChnology) make transputer-based parallel
systems ranging from single processor plug-in boards for IBM PCs through
to stand-alone systems with thousands of processors. The company was
founded by Falk-D. Kubler, Dr.Gerhard H. Peise, and Bernd Wolff in
Aachen, West Germany in 1985. The company had a turnover on sales of
zero in their first year but this has risen to Dm1.5M in 1986, Dm5.2~M
in 1988, Dm9M in 1989, and Dm15M in 1990. They initially targetted
their systems at industry and have a large export market, which in 1988
accounted for one third of their turnover. By mid 1991 the 130-strong
workforce realised a worldwide turnover of around $17~million.
The company ParaCom was formed to take care of sales and marketing while
Parsytec continued with research and development, which consumes most of
the revenues generated. Parsytec/ParaCom have their headquarters in
Aachen, Germany and subsidiary offices in Chicago-USA, Southampton-UK,
and Chemnitz-Germany.
Parsytec's product range includes a mid-range system and two large
stand-alone parallel systems:
MultiCluster (T800 processors).
SuperCluster (T800 processors) --- SC-1/16--64.
Parsytec GC (T9000 processors) --- GC-1 through to GC-5.
Since the T9000 processor is not yet in production the SuperCluster
systems based on the T800 processor are available under the ``Parcour
Program'' as application development systems. Upgrading from the SC to
the GC is claimed to be straightforward since the T9000 is the Inmos
successor to the T800 and has the same instruction set (they also have a
very similar compute/communication performance ratio). Applications
which run on the SC platform are claimed to run 10 times faster on an
equivalent sized GC system.
Platforms Documented:
Contact Address:
See Also:
The GPMIMD Interview. October/November 1991 Parallelogram.
Parsytec MultiCluster (T800
processors).
Overview of Platform:
Compute Hardware:
Interconnect / Communications System:
Memory System:
Benchmarks / Compute and data transfer performance:
Operating System Software and Environment:
Networkability/ I/O System / Integrability / Reliability / Scalability:
Notable Applications / Customers / Market Sectors:
Overall Comments:
Parsytek SuperCluster (T800 processors) SC-1/16--64.
Overview of Platform:
Compute Hardware:
Interconnect / Communications System:
Memory System:
Benchmarks / Compute and data transfer performance:
Operating System Software and Environment:
Networkability/ I/O System / Integrability / Reliability / Scalability:
Notable Applications / Customers / Market Sectors:
Overall Comments:
Parsytec GC (T9000 processors) --- GC-1 through to GC-5.
Overview of Platform:
The GC T9000 machine
- Architecture --- A GC machine can be built up from a number of
GigaCubes. Each GigaCube represents a self-contained unit with its own
power supply, I/O channels and interconnection to other GigaCubes. Each
GigaCube contains 64 T9000 processors packaged at high-density.
Machines larger than the GC-3 (>= 1024 processors) require water
cooling which is facilitated by the use of ``heat pipes''. This allows
a peak performance density of more than 9 GFLOPS/(m3).
- Node --- A GigaCube consists of four clusters of
16 processors and has self-contained redundancy, control processor,
power supply and cooling. A cluster is the basic architectural unit and
consists of 16 Inmos IMS T9000 transputers running at 50~MHz, the
EDC-protected memories (up to 32~Mbytes per T9000), a further redundant
T9000, the local link connections and 4 Inmos C104 routing chips. Each
link of the T9000 is connected to a different C104, thus making it
hardware fault tolerant. Redundancy in a cluster ensures overall
probability of failure is less than that of a single typical chip. A
GigaCube can be purchased as a stand-alone machine, the GC-1/64.
- I/O --- Inside a GigaCube each processor cluster has eight
dedicated links with a bidirectional bandwidth of 160 Mbytes/s. Each of
the two sets of 16 links with an additional control link forms a basic
I/O channel. These are logically driven by the control processor and
therefore allow it to control the attached devices if required. For the
largest systems shared I/O devices amongst the GigaCubes is achieved
with a special module (IONM) which may be cascaded. A general purpose
high speed interface is available called MSC-2 which interfaces
four T9000 links to four SCSI-2 buses, and can provide 20~Mbytes/s per
link. A HiPPI interface is also available with a unidirectional
bandwidth of 100 Mbytes/s.
- Topology --- The communications structure of the machine is
software configurable. Each T9000 has four hard links and up to 16384
virtual links. Each hard link is connected to a C104 32x32 way
cross-bar switch. The C104 can determine the destination of a message
and switch automatically with extremely low latency. Some example
topologies are 2 or 4 dimensional arrays, trees and hypercubes.
- Operating System --- PARIX based on UNIX with parallel extensions,
supports Remote Procedure Call. The I/O library is a subset of the
POSIX standard.
- Communication Paradigms --- GMD Argonne macros for FORTRAN (PARMACS).
Can also support iPSC, Suprenum, nCUBE, Helios-CDL primitives.
- Languages ---
ACE compilers are available for C, ANSI-C, Fortran-77, Pascal and
Modula-2.
- Programming Environment --- Source level debugging. Monitoring and
profiling available using PATOP.
- Performance ---
For the GC-1/64:
- Peak performance = 12.8 GIPS (32-bit), 1.6 GFLOPS (64-bit).
For the GC-5 system with 16384 processors:
- Peak performance = 3200 GIPS (32-bit), 400 GFLOPS (64-bit).
- Sustained peformance = 190 GFLOPS (64-bit).
- Data Transfer ---
Communication bandwidth of a GigaCube:
- Internally = 1.3 Gbytes/s.
- Between processors = 5.1 Gbytes/s.
- Between clusters = 160 Mbytes/s.
- Between GigaCubes = 1.1 Gbytes/s.
- Between arbitrary processors = 80 Mbytes/s.
- I/O bandwidth of a GigaCube:
- GC-1/64 = 0.64 Gbytes/s.
- GC-5/16384 is maximally 160 Gbytes/s.
- Balance --- The communication bandwidth to compute ratio is
claimed to be very similar to the T800. Therefore the same range of
applications that are performed by the T800 systems should be applicable
to the GC T9000.
- Scalability --- The Parsytec GC machine scales from 64 (GC-1/64)
to 16384 (GC-5) processors. The smallest increment is a GigaCube
containing 64 processors. Therefore the largest machine consists of 256
GigaCubes.
- Fault Tolerance --- The GC machine is designed to be fault
tolerant. Within each node there are a number of redundant processors
which can be switched in automatically to replace failed units. Fault
tolerant hardware can handle serious errors such as memory failure.
System Management Software handles single processor failure and provides
advice to the system manager for the location of faults. Applications
can also be restarted from checkpoints, where the context of the program
is periodically saved.
- Price Performance --- $300,000 per GFLOP on entry level systems.
User Base
Since 1986, more than 700 Parsytec parallel computers have been
installed world-wide, being used for compute-intensive applications and
industrial realtime systems.
In 1991 the National Center for Fluid Simulation at the RWTH Aachen was
equipped with a large Parsytec parallel computer funded by the DFG
(German Research Association). The PC2 (Paderborn Center for
Parallel Computing) has installed the largest freely-configurable
parallel computer in Europe. The Institute for Inter-disciplinary
Scientific Computing (IWR) in Heidelberg has been working with a
Parsytec computer as their central system since 1989. Shell in
Amsterdam carries out application research with the largest industrial
parallel computing system based on the transputer.
Some other fields of application are research into parallel algorithms,
neural networks, fuzzy logic and simulation.
Additional Information
Parsytec were involved in the first phase of the GPMIMD project.
However, disagreement with the other members, Meiko, Parsys, Inmos and
Telmat, on the adoption of a single physical architecture, prompted them
to announce their own T9000 machine the ``Grand Challenge'' system as
previously described.
Compute Hardware:
Interconnect / Communications System:
Memory System:
Benchmarks / Compute and data transfer performance:
Operating System Software and Environment:
Networkability/ I/O System / Integrability / Reliability / Scalability:
Notable Applications / Customers / Market Sectors:
Overall Comments:
hawick@npac.syr.edu
saleh@npac.syr.edu