A Study on the Impact of Memory Consistency Models on Parallel Algorithms for Shared-Memory Multiprocessors

Memory consistency model is an integral part of the shared-memory multiprocessor system, and directly affects the performance. Most current multiprocessors adopt relaxed consistency models in quest of higher performance. In this paper we study the impact of memory consistency model on the design, implementation and performance of parallel algorithms for graph problems that remain challenging due to poor locality. Our study reveals interesting findings through a unified algorithmic and architectural approach. We show that relaxed models do not necessarily complicate algorithm design, and the simplicity and intuitive ordering constraints of sequential consistency can help reduce synchronization and achieve superior performance over algorithms for relaxed models. The graph algorithms we study constitute a new benchmark suite for evaluating the consistency model and other aspects of shared-memory systems. Our benchmark suite represents combinatoric workload and lays higher stress on the memory sub-system, and hence complements the SPLASH/SPLAH-2 suites for shared-memory multiprocessors.