Library Cache Coherence

Abstract

Directory-based cache coherence is a popular mechanism for chip multiprocessors and multicores. The directory protocol, however, requires multicast for invalidation messages and the collection of acknowledgement messages, which can be expensive in terms of latency and network traffic. Furthermore, the size of the directory increases with the number of cores. We present Library Cache Coherence (LCC), which requires neither broadcast/multicast for invalidations nor waiting for invalidation acknowledgements. A library is a set of timestamps that are used to auto-invalidate shared cache lines, and delay writes on the lines until all shared copies expire. The size of library is independent of the number of cores. By removing the complex invalidation process of directory-based cache coherence protocols, LCC generates fewer network messages. At the same time, LCC also allows reads on a cache block to take place while a write to the block is being delayed, without breaking sequential consistency. As a result, LCC has 1.85X less average memory latency than a MESI directory-based protocol on our set of benchmarks, even with a simple timestamp choosing algorithm; moreover, our experimental results on LCC with an ideal timestamp scheme (though not implementable) show the potential of further improvement for LCC with more sophisticated timestamp schemes.