Workload Characterization and Locality Management for Coarse-grain Multiprocessors 1 Workload Characterization and Locality Management for Coarse-grain Multiprocessors

Scalable shared memory multiprocessors commonly employ replication and the associated coherency maintenance of memory blocks, but diier in the granularity from ne-grain (cache-coherent multiproces-sors) to coarse-grain (page-based distributed shared memory systems). Regardless of the size of coherency blocks, attaining good performance may depend on the number of copies staying small. Previous work-load characterization studies of ne-grain systems have shown sharing patterns that tend toward small worker sets (set of processors sharing a block). Coarse-grain architectures ooer scalability advantages unless the sharing patterns for larger coherency blocks are considerably less favorable. This is the question we address in this paper. Our results are encouraging in that, despite the coarse granularity, worker sets scale reasonably.