Shared vs. snoop: Evaluation of cache structure for single-chip multiprocessors

The shared cache structures and snoop cache structures for single-chip multiprocessors are evaluated and compared using an instruction level simulator. Simulation results show that 1-port large shared cache achieves the best performance if there is no delay penalty for arbitration and accessing the bus. However, if 1-clock delay is assumed for accessing the shared cache, a snoop cache with internal wide bus and invalidate style NewKeio protocol overcomes shared caches.