Snoop-Free Multicore Architecture based on Dual-Core Clusters

The traditional multicore architecture provides a private cache for each processor and includes a dedicated unit to maintain data coherence among them [1], which degrades the overall performance significantly as it frequently prevents the processor from accessing the cache. A shared cache system has been presented to reduce the burden of coherence management [2][3]. As the number of processors increases, however, the efficiency of cache sharing diminishes due to coincident accesses. We propose and implement a new multicore architecture that shares single-port caches in a cluster consisting of two processors and provides a massage-passing communication to maintain global consistence among clusters.