Processor scheduling for transputer networks

• Main Author: Laghari, M. S.
• Published: Swansea University 1993
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.637839

There are many factors which affect the performance of message-passing parallel processing systems. These include: procesSor scheduling, network topology and the particular decomposition of an application to exhibit parallelism. Applications are commonly decomposed into processes by using three paradigms of parallel processing which are algorithmic and geometric parallelism, and processor farming, to exploit the parallelism inherent in the application. Static and dynamic processor scheduling techniques are used to assign these processes to specific processors. The processors are then interconnected in a network with a topology best suited for the particular application. The work presented in this thesis investigates the performance of scheduling techniques and problem decompositions for the parallel implementation of geometric and grid applications which may be best suited for SIMD implementation. However, by using the geometric and processor farming paradigms, the MIMD structure of transputers is exploited for these applications. The aim of the work is to find parameters for optimum performance. Three application areas are investigated for parallel implementation. These are: <i>cellular automaton</i> simulation, the <i>Hough transform</i> for line detection, and <i>wear particle</i> identification and classification using computer vision. Experiments are performed using different scheduling schemes and problem decompositions for all three applications. Results on transputer networks are obtained by varying the network sizes, computation loads and size of the work packets. The results are compared in terms of total processing times, speedup and efficiency of the parallel processing system to find optimum performance parameters. The experiments demonstrate that transputer networks can be successfully applied to the parallel implementation of geometric and grid-type applications. The dynamic scheduling scheme seems to provide the best performance. However, the optimum performance parameters would critically depend on the size and nature of the particular application.