Research And Implementation Of Caching Techniques On Multi-application Shared Key-value Stores

LSM-tree-based Key-Value stores satisfy the storage scalability demands for the evergrowing data volumes produced by internet services,while,on the one hand,cache resources have always been scarce and are still performance-critical to almost all storage systems.On the other hand,multiple internet services,such as mobile transportation,food delivery,and office share Key-Value store resources and present periodical workload patterns,which challenges LSM-tree-based Key-Value store cache designs.To overcome these challenges,this paper proposes Concerto,which is a tiered cache that hybrids the latest Intel Optane SSD(Solid State Drive)with DRAMs and proposes a discrete time-aware caching policy named Conductor to improve the cache performance for the forementioned shared cache scenario.Specifically,considering that workloads of multiapplication shared Key-Value stores shift periodically,Conductor slices everyday workloads into discrete fixed-size time shards and identifies frequently-accessed hot data for each discrete time shards of the past N days,then prefetches hot data for the upcoming time shard based on its corresponding past time shards from past N days.Besides,Concerto proposes an efficient hash index for retrieving and updating.For efficent cache space managements,Concerto takes advantage of robust performance of the Optane SSD and proposes a lightweight and memory-efficent management for varied Key-Value sizes.Based on LevelDB,this paper prototypes Concerto and evaluates it with periodical workloads.Evaluation results show that Concerto improves the cache hit ratio by up to 11%,decreases the average request latency by up to 13%,and reduces the amount of migration data by up to 92%.