Abstract
This paper proposes and evaluates Sharing/Timing Adaptive Push (STAP), a dynamic scheme for preemptively sending data from producers to consumers to minimize criticalpath communication latency. STAP uses small hardware buffers to dynamically detect sharing patterns and timing requirements. The scheme applies to both intra-node and inter-socket directorybased shared memory networks. We integrate STAP into a MOESI cache-coherence protocol using heuristics to detect different data sharing patterns, including broadcasts, producer/consumer, and migratory-data sharing. Using 12 benchmarks from the PARSEC and SPLASH-2 suites in 3 different configurations, we show that our scheme significantly reduces communication latency in NUMA systems and achieves an average of 10% performance improvement (up to 46%), with at most 2% on-chip storage overhead. When combined with existing prefetch schemes, STAP either outperforms prefetching or combines with prefetching for improved performance (up to 15% extra) in most cases.
Date of this Version
11-12-2013