Silkstore:Write-optimized And Workload Adaptive Key-Value Store

Key-value storage systems,due to its simple data model,have been used widely in social networks,e-commerce,and cloud infrastructures.One important type of such system is based on the idea of Log-Structured-Merge-Tree.Such systems,with excellent write performance,have been playing dominant roles in write-intensive applications.However,there still exists the problem of serious write amplication in those storage systems.That is,the amount of data written to the storage device is much larger than users requested.Write amplication wastes device I/O throughput and wears out devices more quickly.There are academic studies,such as WiscKey,PebblesDB,LSM-Trie,on reducing write amplification.However,such studies ususally sacarfice the performance of reads for writes.Therefore,this paper set out to study this problem with a goal of reducing write amplification without comprimising read perforamnce.This work proposes SilkStore,a write-efficient and adaptive key-value store.This work makes the observation that the multi-level design of traditional LSMTs and compaction algorithms are the main contributers of write amplication.Therefore,SilkStore address this by introducing a single level LSMT design and a more efficient compaction procedure.This work also gives a theoretical analysis to show the advanteges and disadvantages of SilkStore compares to other LSMT designs.Besides,existing LSMT designs are not optimized for workloads that have different read and write patterns to different sub key-ranges.To addres this problem,SilkStore proposes a technique to adaptively optimize the physical organizations of sub key-ranges of the single level LSMT based on read and write hotness measures.This technique shaves off latencies effectively in a worlkoad-aware way.Our experiments show that SilkStore reduces write amplification by a factor of up to 5.1x compare to state-of-the-are solutions,while still maintains excellent point read and range read performance.Furthermore,the adaptive optimization technique effectively finds optimization candidates and reduces read latency by another 27-37% in our experiments.