A timestamp-based selective invalidation scheme for multiprocessor cache coherence

Among all software cache coherence strategies, the ones that are based on the concept of timestamps show the greatest potential in terms of cache performance. The early timestamp methods suffer from high hardware overhead. Improvements have been proposed to reduce hardware overhead at the expense of either increasing runtime overhead or sacrificing cache performance. We discuss the limitations of the previous timestamp-based methods and propose a new software cache coherence scheme. Our scheme exploits the inter-level locality with significantly less hardware support than the early timestamp methods while introducing only constant runtime overhead for each epoch during the execution of a program. Simulation results show that the proposed scheme achieves higher performance than the previous schemes with comparable hardware overhead.