Are distributed sharing codes a solution to the scalability problem of coherence directories in manycores? An evaluation study

The development of efficient and scalable cache coherence protocols is a key aspect in the design of manycore chip multiprocessors. In this work, we present a comprehensive evaluation of a kind of cache coherence protocols that, despite having been already implemented during the 1990s for building large-scale commodity multiprocessors, have not been considered in the context of chip multiprocessors yet. In particular, we evaluate two directory-based cache coherence protocols based on the idea of having the sharing code of each memory block distributed between the different sharers (distributed sharing code). The first one employs simply-linked lists to encode the information about the sharers of the memory blocks, whilst the second one does the same using doubly-linked lists, which improves the management of replacements. We compare these two organizations with three protocols that use centralized sharing codes, each one having different directory memory overhead: one of them implementing a non-scalable bit-vector sharing code and the other two implementing more scalable limited-pointer schemes with one and two pointers, respectively. Simulation results show that for large-scale chip multiprocessors, the protocol based on distributed doubly-linked lists dramatically reduces the memory overhead of a non-scalable bit-vector directory, while at the same time it achieves its performance levels. This is achieved with just a small degradation on dynamic energy consumption (approximately 10 % on average). This way, our results point out that for manycores, coherence directories based on distributed sharing codes are appealing alternatives to contemporary coherence directories based on centralized sharing codes.