PERFORMANCE STUDIES OF PARALLEL REAL-TIME CONTROLLERS

Abstract

This paper first describes performance evaluations tools which enable the control engineer to select appropriate algorithm realisations and hardware configurations for a transputer-based implementation using the occam programming language. An environment has been developed, which integrates the control system design tool, MATLAB, with the Transputer Development System, TDS, plus software tools to automate the mapping of controllers on transputer arrays, using either static or dynamic task allocation. Performance evaluation is carried out on the actual transputer hardware, by displaying on-line, processor activity, execution time and task allocation for each array. The paper further analyses the performance benefits of dynamic and static task allocation, for implementing real-time controllers. A Case Study algorithm has been partitioned into a number of fixed-size small grains, providing an initial insight into the potential benefits and bottlenecks of the allocation strategies. Then, the complexity of the tasks has been augmented, by means of increasing the granularity. This permits the evaluation of the performance of the allocation tools in a more general manner. Further, the tools assist the identification of an appropriate task grain size to result in a suitable compute/communicate ratio for the target hardware.