An effective full-map directory scheme for the sectored caches

In multiprocessor systems, the cache misses due to coherence transactions make up many of the total cache misses. However this type of cache miss is strongly dependent on the type of data sharing among processors, especially false sharing. Until now the small cache block size has been used to avoid false sharing mainly in multiprocessor systems, but the smaller the cache block size, the lower the prefetching effect. Moreover it is shown that high spatial locality appears in many parallel programs. The paper presents two advanced full-map directory schemes which provide a low cache miss ratio and communication traffic by avoiding false sharing and taking advantage of the spatial locality existing in many parallel programs. The performance was evaluated by the event-driven simulator and the empirical results show that the proposed scheme can provide about a 6/spl sim/77% decrease in the cache miss ratio and a 46/spl sim/96% decrease in the communication traffic.