Origin 2000 design enhancements for communication intensive applications

The SGI Origin 2000 is designed to support a wide range of applications and has low local and remote memory latencies. However it often has a high ratio of remote to local misses. In this paper, we evaluate the Origin 2000 performance with communication intensive applications. We use detailed execution-driven simulation of six shared-memory applications. This paper evaluates a base, Origin 2000-like system and three derived systems that incorporate techniques to reduce the communication cost by lowering the ratio of remote misses. We show that the performance of these applications is generally improved when the local bus is used in the snoopy mode, the number of processors per node is increased, the processors use the Illinois cache coherence protocol, or when adding a snoopy cache to retain remote data within each node. Illinois protocol and the interconnect cache reduce the average remote miss ratio by 16% and 21%, respectively.