Research And Implementation Of A Key-value Storage System Based On NVM

Non-Volatile Memories(NVMs)have characteristics of low latency,high storage density,durability and byte-addressability,and they are now publicly available and are promising supplementations of traditional memory devices for next-generation storage systems.Existing key-value store index methods are dominated by Log-structured Merge tree(LSM-tree),while this structure shows suboptimal performance in NVMs due to its background compaction operations and hierarchical structure.It is critical to efficiently integrate NVMs into key-value stores as they are widely proven to be the proper infrastructure for large scale storage systems.Aiming at how to build an NVM-based efficient key-value store system,A highperformence key-value store which named HBKV(Hybrid B+tree Key-Value store)is proposed.HBKV separates keys and values and integrates a carefully crafted DRAM-NVM hybrid index structure,termed HBTree(Hybrid B+tree).HBTree consists of three layers:the indexing layer,the middle layer,and the data layer.The indexing layer is built on a B+tree and HBTree stores this layer in DRAM for the better performance,as DRAM outperforms NVM and this layer is frequently accessed.The middle layer is used to connect the indexing layer and the data layer,and this layer also helps sieve the hot objects in the data layer.The data layer has many small Log Trees,and HBTree caches hot Log Trees in DRAM to improve the read performance.HBTree uses logs to ensure the data reliability of write requests to cached Log Trees.The proposed HBTree indexing structure makes HBKV efficiently access and persist data in NVMs.HBTree and HBKV are prototyped on the Intel Optane DC Persistent Memory produce.HBKV is compared with NVM-based B+tree methods FAST&FAIR and FPTree as well as LSM-Tree methods Level DB and Rocks DB.Evaluations results with the benchmark YCSB show that,HBTree is 1.2 to 1.7 times faster than FAST&FAIR and is 1.1 to 1.8 times faster than FPTree.HBTree also reduces the recovery time by 70% when compared to FPTree.As for comparisons to LSM-Tree methods,HBKV is 3.9 to 12.8 times faster than Level DB and1.9 to 6.2 times faster than Rocks DB.