Optimization of a Linked Cache Coherence Protocol for Scalable Manycore Coherence

Despite having been quite popular during the 1990s because of their important advantages, linked cache coherence protocols have gone completely unnoticed in the multicore wave. In this work we bring them in the spotlight, demonstrating that they are a good alternative to other solutions being proposed nowadays. In particular, we consider in this work the case for a simply-linked list-based cache coherence protocol and propose two techniques, namely Concurrent Replacements (CR) and Opportunistic Replacements (OR), aimed at palliating the negative effects of replacements of clean data. Through detailed simulations of several SPLASH-2 and PARSEC applications, we demonstrate that, armed with CR and OR, simply-linked list-based protocols are able to offer the performance of a non-scalable bit-vector directory at the same time that scalability to larger core counts is preserved.