Abstract

This thesis evaluates directory-based cache-coherence protocols for COMA machines which are an emerging class of large-scale shared-memory multiprocessors. The cache-coherence protocol is a critical issue in designing such machines because it directly affects memory latencies. Single-level full-map directory protocols can suffer from contention in a large-scale machine. To alleviate this contention and increase concurrency, hierarchical directory protocols may be used. By simulating with parallel applications from the SPLASH benchmark suite, we compare the performance, scalability, and resource usage of a single-level scheme with those of a hierarchical protocol. Our results show that performance depends on the data access patterns and degree of synchronization in the program. The full-map scheme performs adequately with a low processor count or for programs with low synchronization. However, the hierarchical schemes scale better and perform well as the number of processors and the synchronization overhead increases. Good locality in a program also makes a hierarchical protocol more attractive, reducing miss latencies and required network bandwidth. Before picking a protocol, a system architect should carefully balance performance, memory overhead, and network bandwidth considerations.

Citations