Optimizing Private Memory Performance by Dynamically Deactivating Cache Coherence

User Program has significant percentage of memory requests that target only private data, which does not need to resolve cache coherence conflicts. Yet traditional coherence protocol does not distinguish between shared and private blocks, which leaves much optimization space for memory reference latency and power consumption. This paper proposes an optimized cache coherence protocol, called PMESI, which dynamically deactivates cache coherence maintenance for private memory space. PMESI achieves two distinguishing features: the reduction of memory access latency and system power consumption. Simulation results on the cycle accurate simulator for wide range of industry benchmarks show that 54% memory references can be efficiently optimized and program execution time is reduced 9% on average.