SCIP: Selective Cache Insertion and Bypassing to Improve the Performance of Last-Level Caches

The design of an effective last-level cache (LLC) is crucial to the overall processor performance and, consequently, continues to be the center of substantial research. Unfortunately, LLCs in modern high-performance processors are not used efficiently. One major problem suffered by LLCs is their low hit rates caused by the large fraction of cache blocks that do not get re-accessed after being brought into the LLC following a cache miss. These blocks do not contribute any cache hits and usually induce cache pollution and thrashing. Cache bypassing presents an effective solution to this problem. Cache blocks that are predicted not to be accessed while residing in the cache are not inserted into the LLC following a miss, instead they bypass the LLC and are only inserted in the higher cache levels. This paper presents a simple, low-hardware overhead, yet effective, cache bypassing algorithm that dynamically chooses which blocks to insert into the LLC and which to bypass it following a miss based on past access/bypass patterns. Our proposed algorithm is thoroughly evaluated using a detailed simulation environment where its effectiveness, performance-improvement capabilities, and robustness are demonstrated. Moreover, it is shown to outperform the state-of-the-art cache bypassing algorithm in both a uniprocessor and a multi-core processor settings.