Evaluation of Low-Overhead Organizations for the Directory in Future Many-Core CMPs

Abstract. If current trends continue, today’s small-scale general-purpose CMPs will soon be replaced by multi-core architectures integrating tens or even hundreds of cores on-chip. Most likely, some of these many-core CMPs will implement the hardware-managed, implicitly-addressed, coherent caches memory model. Cache coherence in these designs will be probably maintained through a directory-based cache coherence protocol imple-mented in hardware. The organization of the directory structure will be a key design point due to the requirements in area that it will pose. In this work, we study the effects on performance, network traffic and area that the use of compressed sharing codes for the directory will have in many-core CMPs. In particular, we select two compressed sharing codes previously proposed in the context of large-scale shared-memory multiprocessors that have very small area requirements. Simulation results of 32-core CMPs show that degradations of up to 32 % in performance and 350 % in network traffic are experienced. Additionally, since some proposals for efficient multicast support in on-chip networks have recently appeared, we also consider the case of using this support in combination with the compressed sharing codes. Unfortunately, we found that multicast support is not enough to remove all the performance degradation introduced by the compressed sharing codes and barely can reduce network traffic.