| The emerging Non-Volatile Memory(NVM)has advantages of high performance,byte-addressability,and non-volatility.NVM is usually used with the Solid State Drive(SSD)to balance cost and performance.The key-value store technology based on the NVMSSD hybrid structure has become an important research direction.The Log-Structured Merge Tree(LSM-tree)is widely used in key-value stores.However,the read and write amplification and the write stalls caused by the hierarchical structure can harm the system performance.Existing studies based on hybrid storage only focus on reducing write stalls or reducing read amplification at the expense of data ordering in SSDs,which cannot improve read performance,write performance,and scan performance at the same time.It is necessary to do further research on high-performance key-value stores.To improve read performance,write performance,and scan performance in the NVMSSD hybrid LSM-tree,Hybrid Plane Key-Value(HPKV)is proposed.HPKV is a plane LSM-Tree key-value store based on the NVM-SSD hybrid structure.First of all,a plane three-level hybrid dynamic partition structure is adopted in HPKV.Each partition includes three parts: an upper DRAM index,a middle NVM cache,and a single-level LSM-tree.The DRAM index and the NVM cache are used to avoid log overhead and flush data from NVM to SSD efficiently to reduce write stalls.The operation of searching the level LSM-tree is accelerated by an NVM B+-tree to reduce read amplification.A dynamic partition strategy is adopted to improve the scan performance of the single-level LSM-Tree.Secondly,to improve read performance,a double index structure combining hash-table and B+-tree in DRAM is used for data in the NVM cache.Finally,a batch-update strategy is adopted to reduce the maintenance overhead of the persistent B+-Tree and improve the efficiency of flush operation and garbage collection.The prototype of HPKV is implemented in the NVM-SSD hybrid environment.The performance of HPKV is evaluated to compare with popular LSM-Tree key-value stores.Test results show that HPKV performs well under various YCSB(Yahoo! Cloud Serving Benchmark)workloads.The read performance of HPKV is about 1.1 times higher than that of Nove LSM,and 1.7 times higher than that of Matrix KV.The write performance of HPKV is about 1.2 times higher than that of SLM-DB.The scan performance of HPKV is about 0.7 times higher than that of SLM-DB. |
