A Quantitative Performance Evaluation Of Sci Memory Hierarchies

The simplest topology supported by SCI is the ring. It was found that, for the hardware and software simulated, the largest efficient ring size is between eight and sixteen nodes and that raw network bandwidth seen by processing elements is limited at about 80Mbytes/s. This is because the network saturates when link traffic reaches 600- 7000Mbytes/s. These levels of link traffic only occur for two poorly designed programs. The other four programs generate low traffic and their execution speed is not limited by interconnect nor cache coherence protocol. An analytical model of the multiprocessor is used to assess the cost of some frequently occurring cache coherence protocol operations. In order to build large systems, networks more sophisticated than rings must be used. The performance of SCI meshes and cubes is evaluated for systems of up to 64 nodes. As with rings, processor throughput is also limited by link traffic for the same two poorly designed programs