| The cache only memory architecture, COMA, exploits data locality by supporting automatic data migration in hardware but suffers from long communication latencies. Multithreading is a latency hiding technique that improves the processor utilization by activating a new thread while other threads are waiting on events. By combining COMA and multithreading, multithreading will benefits from the dynamic data migration of the COMA and, at the same time, it will hide long communication latencies of COMA which will improve the overall system performance.;In this thesis, we describe new cache only memory architecture, MCOMA, whose performance supersedes any of the existing COMA architecture. The system consists of a number of nodes connected by a mesh interconnection network to move data between nodes. Each group of nodes has a group directory. The group directories are connected by a separate search interconnection to improve the data search.;To improve the performance further, the new architecture is combined with multithreading. This architecture benefits from the data locality of COMA and the latency tolerance technique of multithreading. We developed a simulator to evaluate and study the performance of the new system. The simulator is execution driven, which is developed on top of MINT, a MIPS interpreter. We evaluated the performance of the architecture using a subset of SPLASH benchmark suite of parallel application programs covering different problem domains.;The performance evaluation has been done for a wide variation of architecture parameters. The effect of multithreading on execution time, cache hit rates, attraction memory hit rates and group hit rates of the system have been evaluated. The effect of different cache sizes, as well as, the memory block length on system performance has also been evaluated, analyzed and presented. The simulation results show MCOMA has achieved a good overall speedup on all the used applications. All the applications performance improved by the use of multithreading on the MCOMA which proved that combining the benefits of dynamic data migration in COMA with multithreading improved the overall system performance. |
