Sensitivity Analysis of Transputer Workfarm Topologies
Timothy J. Johnson · Calhoun: The Naval Postgraduate School Institutional Archive (Naval Postgraduate School) · 1989
Parallel processing structures such as multiprocessor arrays and pipelining enhance throughput tremendously for suitable algorithms having high degrees of concurrency. However, if the time to process different workpackets becomes irregular, much of the advantage offer traditional sequential processing systems may be lost. In an attempt to produce a more flexible response to workload demands, a transputer workfarm was investigated. Two network topologies, a linear model and a tree model, were built using the transputer as the processing element (PE) or worker. An algorithm was developed which could be run independently on all workers in the workfarm. Each worker produced results independent of the other workers. By altering specific variables within the algorithm, the network performance could be changed. The results from this thesis illustrate how these parameters affect each network and provide comparative information between the linear model and the tree model.