A Simple and Efficient Cache Coherence Protocol Based on Self-Updating

As the number of cores in a chip multiprocessor increases, cache coherence protocols have become a performance bottleneck of the share-memory system. The overhead and complexity of current cache coherence protocols seriously restrict the development of the share-memory system. Specifically, directory protocols need high storage overhead to keep track of sharer list and snooping protocols consume significant network bandwidth to broadcast messages. Some coherence protocols, such as MESI (modified exclusive shared or invalid) protocol, are extremely complex and have numerous transient states and data race. This paper implements a simple and efficient cache coherence protocol named VISU (valid/invalid states based on self-updating) for data-race-free programs. VISU is based on a self-updating mechanism and only includes two stable states (valid and invalid). Furthermore, the VISU protocol eliminates the directory and indirection transactions and reduces significant overheads. First, we propose self-updating shared blocks at synchronization points for correction with the data-race-free guarantee of parallel programming. Second, taking advantage of techniques that dynamically classify private data (only accessed by one processor) and shared data, we propose write-back for private data and write-through for shared data. For private data, a simple write-back policy can reduce the unnecessary on-chip network traffic. In L1 cache, a write-through policy for shared data which can keep the newest shared data in LLC, would obviate almost all coherence states. Our approach implements a truly cost-less two-state coherence protocol. The VISU protocol does not require directory or indirect transfer and is easier to verify while at the same time obtains similar even better performance of MESI directory protocol.