Research On LSM-tree Based Key-value Store On Open-channel SSD Features

With the development of large-capacity Solid State Drive(SSD)with big blocks and 3D NAND Flash technology,data storage management is facing some opportunities and challenges.Log-structured merge-tree(LSM-tree)based key-value(KV)store is widely used as a persistent storage engine in data center,since it can optimize for random writes and provide efficient data insertion and deletion.LSM-tree-based KV store usually uses SSD block devices.However,both of the LSM-tree based KV store and SSD block devices have read/write amplification.Even worse,it will be further superimposed because the host and device management are independent.Hence,the performance of the whole storage system and the SSD lifetime will be severely reduced.The LSM-tree based key-value storage system on Open-Channel SSD provides a solution to the unified management of software and hardware resources,optimization of storage systems,and reduction of read/write amplification.We design and implement a LSM-tree based key-value Store System on Open-Channel SSD Features(LOSF).Open-Channel SSD awares semantics of application to classify and aggregate KV data,which selectively separates key and value according to the granularity and access frequency of KV data.It forms small-granular-high-frequency data area managed by LSM-tree structure and large-grained value data area managed by block structure,which improves the efficiency of data recovery and reduces write amplification.For small-grain-high-frequency data,a sub-block management method is proposed to reduce the cost of a single merge operation.A dynamic preprocessing merge strategy is used to reduce the amount of data for each merge operation and effectively reduce the repeated migration of data.The system write amplification is reduced.For large-grained value data,a method for managing the write-in bitmap area is proposed to reduce the cost of LSM-tree compaction,by using rewriting characteristics of the 3D NAND MLC Flash border SLC type media,which reduces read amplification.The test results show that,compared with the LOCS scheme which also utilizing Open-Channel SSD,when the KV granularity is small,the LOSF scheme optimizes the read/write amplification by 30.4% and 29.5%,the average SSD erase cycles reduces by 23.5% and the performance of the storage system increase 28.6%;When the KV data granularity is large,the performance is doubled,and the optimization effect is significant.Compared with the Wisckey-SSD scheme which the host side uses data classification and aggregation,the LOSF scheme optimizes read/write amplification by 35% and 45%,reduces the number of SSD erase cycles by about 70%,and improves storage system performance by 47%.