COMPUTERS AND MORE

Array Architectures

As examples of SIMD array processors, the ICL Distributed Array Processor (DAP) and the TMC Connection Machine have particular significance by virtue of their position in the chronology and taxonomy of these types of machine. The DAP, for example, was the first commercial exploitation of this style of architecture, a style which can be traced back many years. In particular, the design of the DAP owes much to the pioneering work carried out on the SOLOMON computer, and later on the ILLIAC IV computer. The Connection Machine represents a more recent evolutionary step embodying the use of several processing elements on a single chip, and the consequent production of a massively parallel system.

Multi-processors and Multi-computers

Most computer scientists, and users of MIMD machines, draw a distinction between multiprocessor and multicomputer systems. If one considers a processor as simply a component of a computer system, then the distinction becomes clearer. A multiprocessor is then a system in which there is a simple replication of processors within a framework which does not alter the relationship between the processor(s) and other components (such as memory). Conversely, a multicomputer is a system in which the whole computer (processor and memory together) are replicated, and some form of communication network added, to allow them to exchange information. The use of MIMD machines is still in its infancy, and the long-term performance potential of multiprocessor and multicomputer systems is still unclear. One might reasonably ask whether MIMD architectures are really needed, since with hindsight it can be seen that the speed of conventional SISD and SIMD machines increases by an approximately ten-fold factor every five or so years. This has encouraged some users of computers to conjecture that one should simply `wait a few years' for improved technology to provide the required performance. Furthermore, they might argue, by the time a new and novel architecture has been developed it may well be superseded by a faster sequential machine. The problem with this argument is that the development of implementation technologies, sequential processor architectures and MIMD architectures complement each other. Thus, faster sequential processors mean faster processing elements within MIMD machines, and higher performance overall. Ideally, MIMD machines with N processors should simply be N times faster than SISD machines constructed from equivalent technology, although in practice this is rarely the case.