SPECULATIVE INCOHERENT CACHE PROTOCOLS.

This article focuses on speculative incoherent cache protocols. Multiprocessing and multithreading are becoming ubiquitous even on single chips. With increasing cache sizes, coherence misses in such systems will account for a larger fraction of all cache misses. As communication latencies increase, this larger fraction of coherence misses will cause significant and increased performance losses. Tuning coherence proto- cols for specific communication patterns and applications can reduce communication latencies. Speculative execution has successfully improved performance in various scenarios. This article proposes a new type of load speculation, called coherence decoupling. This technique uses speculation to reduce the effect of long communication latencies without exacerbating the programmer's task or complicating the coherence protocol. Separating the SCL protocol and the coherence protocol enables designers to tune and optimize them independently, accelerating communication with less complexity than with conventional protocol optimizations. Separation also lets the two protocols overlap.