Transputer implementation of systolic arrays for model reduction

Abstract

In control theory the dynamics representing a physical system to be controlled is analyzed in terms of a model built up from a mathematical description of the system components. Unfortunately the resulting model can have high or infinite dimensional state equations, so it is often necessary from a practical and computational viewpoint to find a reduced model of the system which mimics closely the behaviour of the high order system. The technique for producing these simplified descriptions is termed model reduction and a number of methods are known. The author is interested in applying computational parallelism to the generation of reduced models to provide an acceleration mechanism for use in a CAD environment. In the paper the author develops systolic algorithms for simple single input-output control problems and implements them on a network of transputers. The performance of this network is then evaluated with respect to a sequential algorithm for the same problems. Finally, the author points out that systolic arrays are normally considered as special purpose methods geared towards exploiting VLSI techniques.