It has long been known that a West German company, Parsytec GmbH of Munich, was playing very seriously with the Inmos International Transputer in the field of industrial controls, but the company has now extended its involvement with the ingenious microprocessor by building a full-scale supercomputer with it. Claiming that the machine paves the way for an era of unprecedented performance levels it has unveiled the Supercluster parallel processor, writes Computerwoche. The theory of vastly improved performance can be achieved on appropriate problems by adopting a parallel processing architecture of course meshes happily with the design of the Transputer. Each Transputer has four built-in high-speed communications links that can connect it to four other Transputers, making it possible to tie autonomous processing nodes together to create a structure that can be adapted to a variety of demanding applications under operating system control. Within the Parsytec Supercluster, each processor uses all four communications channels, and Parsytec argues that not only does the structure solve the bottleneck problems which bedevil traditional bus-based systems, but it confers theoretically unlimited upgrade potential. All current models are based around the 20MHz T800 Transputer, but the company will use the 30MHz T801 version as soon as it is available at the end of the year. But in the present version, with four 20Mbps communications channels and 1Mb of main memory, each node is claimed to achieve a processing rate of 1.5 MFLOPS. The system’s hardware also includes error detection and correction, an inbuilt fault-diagnosis and early warning facility, while plans to develop fault tolerant software – crucial for systems using over 1,000 processors concurrently – are well under way. Also under development are programs to run under the Unix-derived Helios operating system, written in parallel languages such as Occam and Parallel-C: current programs are written in C, Fortran and Pascal. Other features include a Network Configuration Manager to ensure that system processing resources are maximised and distributed to as many simultaneous users as possible.
