FussyCache: A Caching Mechanism for Emerging Storage Hierarchies

In this paper, we propose a novel caching mechanism, called FussyCache, that differs from the traditional DRAM caching mechanisms that automatically cache a data block when it is requested. Instead, FussyCache evaluates each requested data block for its caching eligibility, and reads ineligible blocks directly from the device each time. We show that the Fussy-Cache performs substantially better than the traditional caching algorithms and its performance increases with the storage device speed. In particular, for the first generation Intel Optane based storage, FussyCache provides 25-30% reduction in the average access latency as compared to the native caching mechanism such as plain LRU. We also observe close to 15-20% improvement in performance even for a mainstream TLC SSD. Furthermore, the FussyCache design includes two mechanisms that allow for its easy deployment in any environment: (a) a self-monitoring stage that reverts it to a normal LRU when partial caching is not beneficial, and (b) a training phase that automatically tunes the configurable parameters for the deployment environment.