As reported last week (CI No 1,747), Inmos International Plc, Meiko Scientific Ltd, a spin-off of Inmos, Thorn EMI Plc buy-out Parsys Ltd and Telmat Informatique SA as well as the Centre for European Research into Nucleonics on the Swiss-French border, which is an associate partner, have joined forces to try and develop a standard for parallel processing machines based on the Inmos T9000 Transputer. The Esprit programme under which the standard is to be developed is called GP MIMD, standing for general purpose multiple instruction multiple data. True, as pointed out by the consortium which has been working on the general purpose MIMD architecture for around 18 months, there are too many different parallel architectures around, and it won’t build potential users’ confidence if there isn’t soon some sort of technological compromise. Scientific researchers using old Cray Research Inc systems don’t want to have to re-write all their applications when they realise they need something with more parallelism. What’s needed is a common base architecture, so that programs can be written that will run on parallel systems from different vendors. One of the current limitations that will hamper the construction of a general purpose architecture around the latest generation Transputer, says one observer, is that while the T9000 enables several tasks to share a processor, memory is still limited to the size of physical memory at each processing node – it probably won’t be until the next generation of silicon, expected to be released in 1994, that the Transputer has demand-paged virtual memory. With demand-paged virtual memory, processing sequences can be stopped mid-flow and swapped with another process that is being held on disk – then the original process can be swapped back and resumed at the point where it was put on hold. The introduction of this level of memory management will mean that many applications can be run simultaneously, with very long, time-consuming application projects running continuously in the background.

Adaptive routing

Only then, when the limitations of physical memory are thrown back, will the Transputer be anywhere near forming the base of a truly general purpose parallel machine. The innovative part of the GP MIMD research project is the communication between processors. Time and again parallel supercomputer manufacturers have stressed the importance of data throughput capacity there’s no point having parallel processing power if a bottleneck is going to result from poor communication capacity and flexibility between the processors. The GP MIMD project is using processor communication technology that originated at Southampton University, under Professor Tony Hey and a team of parallel processing experts. The result of this research, which was taken over by Inmos, was the C104 link-switch chip, which when used in conjunction with the T9000 Transputer, enables dynamic communications systems to be built, with adaptive routing. Transputer-based machines of old were based on statically-configured communications systems, which relied on the Transputer’s inherent, but insufficient, message-passing capabilities. With the C104 chip, says Inmos, all machines developed under the GP MIMD project will maintain a balance between computation and global communication. The GP MIMD’s operating system is to be based on Chorus Systemes SA’s microkernel, which will support Unix System Laboratories Inc Unix utilities and System V.4-compatible programming interfaces. – Sue Norris