The decision by ICL Plc and Pyramid Technology Corp to unveil their single image box, tentatively named Merlin, in Australia (CI No 2,504) caused a few eyebrows to be raised – particularly because of the admission by both companies that the appeal of the system is to a fairly limited set of customers. The reason […]
The decision by ICL Plc and Pyramid Technology Corp to unveil their single image box, tentatively named Merlin, in Australia (CI No 2,504) caused a few eyebrows to be raised – particularly because of the admission by both companies that the appeal of the system is to a fairly limited set of customers. The reason for the ploy, it seems, is a question of ownership; ICL is 80% owned by Fujitsu Ltd; Fujitsu manages the Australian business for both and wanted to show the stuff at the Unix ‘Down Under’ Expo, so ICL bowed to the pressure from Tokyo. Merlin, which incorporates ICL’s Goldrush parallel processing system and Pyramid’s Nile servers, integrates symmetric multi-processing and massively parallel operating environment technologies to create a view of the system as a single image, regardless of how many processors are under the hood. Systems management software is wrapped around it. The increasing co-operation between ICL and Pyramid is possible because of the businesses in which each organisation operates. ICL specialises in the public sector, utilities, finance and retail; Pyramid is geared towards the commercial market. The idea behind parallel processing is simply to build big systems from lots of small processor chips. A parallel processor links tens or hundreds of commodity microprocessors into a big system. A massively parallel system can link thousands together. However, for the system to work at anything like its full potential, the work has to be divided, the problem being one of load imbalancing or time and space sharing. There are three ways in which this problem is tackled: divide the problem into horizontal domains which match the architecture; divide a problem into geometric domains; or use temporal (farming) parallelism, in which the problem is a set of independent but equivalent tasks. However, even when the application has been split up, no single part can perform faster than it could as a single chip; hence the move within the industry towards parallel vector processors, systems in which small numbers of processors are running in parallel. Because of these complexities, use of parallel processing has been confined, mainly to academic and research communities. So what is changing? The adoption of low-cost client-server symmetric multiprocessing technology and the take up of relational databases, parallel processor proponents argue, is creating bottlenecks. Shared memory means that the more processors that are added, the worse the bottleneck. The solution is to give each processor its own memory, as in the parallel model. A switch that works like a telephone exchange enables the processors to talk to each other. There are competing versions, however. In the AT&T Corp-Teradata model, processors deal with their own data. In others the disk resources are shared out by a global lock manager and the switch.