Optimal interconnection network bandwidth for ring-based multiprocessor systems

In the past few years, ring-based multiprocessor systems make up the vast majority of small/medium scale shard-memory multiprocessor market. Ring-based system performs better compared to the bus-based system, because ring structure using uni-directional point-to-point links can pipeline the ring traffics. Recently the ring-based system, which sends transactions in opposite directions to enhance the system performance, is introduced. All the introduced systems, regardless of the number of processors, use 1 GB/sec for their ring bandwidth. In this paper we present the optimal interconnection bandwidth for ring-based multiprocessor system changing the number of processors. We use the program-driven simulator to measure the system performance. Simulation results show thar I GB/sec or 2 GB/sec is cost-effective for the system with 8 processors, 2 GB/sec or 4 GB/sec for that with 16 processors, 4 GB for that with 32 processors.