Multilevel Cache Management Based on Application Hints

Multilevel caching is common in many storage configurations, introducing new challenges to cache management. Some of the existing cache replacement policies are designed to improve cache hit rates in the absence of temporal locality, while others avoid replication of data in multiple caches. Additional policies use application hints, an approach that has been shown to significantly improve cache utilization, but only in the highest level of the hierarchy. Our main contribution is a global, non-centralized, dynamic and informed management policy for multiple levels of cache. Karma is aimed at applications, such as databases, which can easily provide hints. These hints are leveraged for making informed allocation and replacement decisions in all cache levels, while preserving exclusive caching, and adjusting to changes in access patterns. This integrated solution is significantly better than other approaches in utilizing the aggregate cache resources. To show the value of Karma, we compare it to LRU, 2Q [17], ARC [22], MultiQ [33], LRU-SP [6] and Demote [32] on database and Zipf traces. For complete evaluation we additionally define extensions of each of these policies to work in a hierarchical cache. On all but very small caches, Karma shows significant improvements over other policies. For example, on a permutation of TPC-H queries, Karma improves over pure LRU by an average of 85%. It adds an average of 50% to the improvement of Demote over LRU and an average of 25% to that of LRU-SP.