Design and Analysis of Reduced-Connection Multiple Bus Systems: A Probabilistic Approach.

Abstract

In the dissertation we have proposed the first systematic and formal approach to reduce connectivity of general purpose multiple bus systems. The approach is based on a probabilistic technique. The hypothesis on which this dissertation is based stipulates that bus connectivity in multiple bus systems can be much reduced by removing connections which are only needed for highly improbable request patterns. With this approach, performance comparable to that of the original multiple bus system could be achieved while significantly reducing memory bus connectivity. The new architecture thus obtained (Probabilistically load reduced connection multiple bus system, or PRMB in short) might have different possible configurations each possibly with a different bus connectivity cost. We have studied the possible relationship among different possible configurations of PRMB systems and proposed an algorithm that determines the one with minimum memory-bus connectivity cost for a given performance level. Our analysis results strongly supported our hypothesis. The queuing problem for PRMB systems is a complicated one because of its unique modeling requirements. An interesting and innovative modification of aggregation technique has been developed to solve queuing problem taking into account bus contention in PRMB system. We have utilized the proposed approximate technique to determine the system throughput. We have also simulated the queuing networks without applying any approximations. Comparison of analytical results with simulation data indicated that our approximate method could accurately be used to model such queuing networks. The results indicated that our hypothesis is valid when queues are utilized. We have proposed another variant of the PRMB system which attempts to reduce the connectivity cost from both the processor side and the memory side. Our results indicated that, except under certain specific conditions, this variant of PRMB system did not offer any cost improvement over the original version. It is quite possible that PRMB system is so efficient that further reduction may not be possible without sacrificing some performance. The technique presented in this dissertation is of very general nature and could possibly be applied to other types of networks as well.