A scalable snoopy coherence scheme on distributed shared-memory multiprocessors

The authors present a scalable snoopy scheme based on a single-hop-connected multiple-bus topology. To reduce the hardware complexity, each processor will snoop on a dynamically changing subset of the buses. The physically-distributed-logically-shared memory model is chosen to let processors take advantage of local memory accesses for private data. The proposed scheme can enjoy fast memory access and can also take advantage of greater scalability. At the same time, the number of transceivers can grow naturally with the number of processors in the system. Comparing this solution with the directory schemes and the traditional snoopy coherent schemes indicates that the performance of the scheme is promising