Study And Implementation On Scalable Hierarchical Cache Coherence Directory Scheme

Distributed shared-memory (DSM) multiprocessors are capable of providing significant performance benefits for scientific and commercial applications. Most recent multiprocessors employ the cache coherent non-uniform memory access (CC-NUMA) architecture. It has important impact on the correctness and performance of DSM systems to maintain cache coherence efficiently and directory protocols have become the first choice to maintain cache coherence of DSM systems because of better scalability. The scalability of directory protocols includes both the memory overhead of directories and their performance, and the memory overhead of directories constitutes a major hurdle in the scalability of CC-NUMA architectures.This paper has proposed a two-level directory organization based on directory cache. The first level directory of this organization is a bit-vector directory organized as a cache and storing the most recently referenced directory lines; and the second level directory is organized as a limited pointers directory which keeps a directory entry for each memory line. This two-level directory architecture reduces the memory overhead of directories and improves the directory information access speed, consequently improving the scalability of directory protocols. We also introduce the design and the realization with FPGA of the directory cache applied in a parallel computer system project.In order to simulate the performance of the system with the two-level directory architecture based on directory cache, we have set up a new simulator after making some necessary revisions on the RSIM simulator explored by Rice University. At last, we run six standard test applications on the simulator to evaluate the effect of the two-level directory on system's performance. We also evaluate the impact of directory cache on the performance of the system under different conditions. The results show that the two-level directory based on directory cache reduces the memory overhead of directory protocols and improves system performance, so it is a method worth further research for improving the scalability of directory protocols.