Application of enhanced JETTY snoop filters to reducing bus traffic and snoop power in SMP and CMP systems

Symmetric multiprocessing systems using snoopy cache coherence protocols dissipate substantial power servicing coherence snoops. The JETTY mechanism was proposed to reduce power use by filtering snoops that the JETTY predicts will miss. This thesis evaluates several improved versions of the JETTY filter, and extends JETTY-style snoop filtering to chip multi-processing (CMP) systems that use snoopy coherence protocols. Practical filters remove 44.8% of snoops in simulated SMP test systems and 52.2% in CMP systems running the SPLASH-2 and TPC-C benchmarks.; JETTY-style filters are also applied to eliminate unnecessary broadcast bus traffic. Filters identify memory accesses known to miss all other processors' caches. These are either marked as non-coherent or diverted directly to memory. Practical filters remove 21.4% of SMP and 41.4% of CMP broadcasts.