A simulation study of region-level coherence tracking in ring-based multiprocessors

This paper presents simulation results on the effectiveness of region-level coherence tracking in a ring-based multiprocessor to complement previous work that implemented coherence tracking for a prototype ring-based multiprocessor in programmable logic. Small memory structures in each node of the multiprocessor distinguish local/remote memory addresses and monitor the sharing of large contiguous regions of memory. The purpose of this monitoring is to filter unnecessary invalidation broadcasts that stem from conventional enforcement of cache coherence. The per-node structures for a larger-scale architecture are adapted from the per-processor filters as proposed for the RegionScout concept. Using a multiprocessor version of the SimpleScalar simulator that has been enhanced with region-level coherence tracking, results for a representative parallel application with writes localized within a node indicate that the per-node tracking structures can filter up to 90% of invalidations exiting or entering a node. When both per-node and per-processors are used, the percentage of invalidation requests that are filtered from exiting nodes is reduced, but the absolute number of invalidations that are permitted to exit nodes is not affected.