Distributed-shared-memory support on the Simultaneous OpticalMultiprocessor Exchange Bus

摘要

This paper examines the performance of distributed-shared-memorynsystems based on the Simultaneous Optical Multiprocessor Exchange Busn(SOME-Bus) using queueing network models and develops theoreticalnresults which predict processor utilization, message latency and othernuseful measures. It also presents simulation results which compare thenperformance of the SOME-Bus, the mesh and the torus. The SOME-Bus is anlow-latency, high-bandwidth, fiber-optic interconnection network whichndirectly links arbitrary pairs of processor nodes without contention. Itncontains a dedicated channel for the data output of each node,neliminating the need for global arbitration and providing bandwidth thatnscales directly with the number of nodes in the system. Each of N nodesnhas an array of receivers, with one receiver dedicated to each nodenoutput channel. The entire N-receiver array can be integrated on ansingle chip at a comparatively minor cost resulting in o(N) complexity.nThe SOME-Bus supports multiple simultaneous broadcasts of messagesnallowing cache consistency protocols to complete much faster. Comparednto the networks considered here, the SOME-bus is the interconnectionnnetwork whose performance is least affected by large messagencommunication times