Performance improvement of the two dimensional broadcast-based distributed shared memory multiprocessor systems by message operations

Distributed shared memory (DSM) systems present strong advantages in terms of scalability, programmability and cost effectiveness. However, obtaining high performance out of these systems turns out to be much more cumbersome than expected due to the increasing remote memory access latency as the number of processors increase. The underlying interconnection network has taken on an increasingly important role in overall performance of DSM systems. Two-dimensional (2D) broadcast-based optical architectures are a popular solution. In this thesis, DSM on 2D broadcast-based optical architecture is simulated by using OPNET Modeler and the performance of the system is improved by message priority and combining techniques under different traffic patterns, including uniform, hot region and bit permutations. The simulated system contains 64 nodes. Each node has a server that processes directory, cache and processor messages, two output channels and two input queues, one for horizontal and the other for vertical dimension. Scenarios are run to measure average network response time, average input waiting time and average channel waiting time for varying values of DSM parameters such as the ratio of the mean channel service time to mean thread run time, probability of a cache block being in modified state and the fraction of write misses. The results show that combining reduces average channel waiting time between 13% and 29%, while priority reduces average input waiting time between 7% and 43% on average. With the application of message priority and combining, between 7% and 40% performance increase for average network response time has gained.