Research Of Monitoring And Filtering Mechanism On Cache Pollution

Most existing CPUs use the least recently used (LRU) algorithm as their cache re-placement algorithm. However, the LRU algorithm cannot effectively identify the localitycharacteristics of the data accessed and timely adjust management strategy according tothe variation of the data’s locality. Specifically, based on processors with shared caches,LRU can cause cache pollution by inserting non-reuseable data into the caches whileevicting reusable ones, resulting in performance degradation. Existing work is lack of anefficient mechanism to deal with such cache pollution problem to avoid performance de g-radation.In this paper, we propose the Monitoring and Filtering Mechanism on Cache Pollu-tion (MFMCP), a dynamic shared cache pollution monitor with standard dynamic memoryallocation interfaces, requiring no special hardware support and no change to upper appli-cations. By integrating a monitor into the memory allocator, MFMCP can identify the lo-cality characteristics of the data accessed and timely adjust management strategy accord-ing to the variationof the data’s locality. The online monitor not only can efficiently mon-itor the cache pollution based on the heap memory, but also can predict the memory typeand then send the predicted memory type to the buddy system with page coloring tec h-nology. Then, the weak-locality data can be dynamically mapped to a small color region inthe cache, eliminating the cache pollution problem.Extensive experiments show that the overhead whic h MFMCP introduced is less than1%, and the MCMCP can improve the performance of memory-intensive SPEC CPU2006benchmarks by20%average.