Research On Write Amplification Optimization Of Key Value Store Based On Open-channel SSD

With the development of key value store based on LSM-tree(Log-structured merge tree)and flash storage technology,the key value storage system using SSD(Solid state drive)as backend storage device has been widely used.However,the information isolation between SSD and the key value store makes the flash device unable to release its potential performance;meanwhile,the write amplification problem caused by the compaction process of LSM tree has adverse effects on the system performance and lifetime of SSD.The emerging OCSSD(Open-channel solid state drive)provides a new solution to these problems.Aiming at the write amplification problem of LSM-tree,combined with the characteristics of OCSSD,this paper proposes a compaction method based on data block remapping,and designs the corresponding key value store,called OCKV.Firstly,to reduce write amplification and improve the efficiency of compaction,OCKV directly remaps the data blocks whose key ranges don't overlap with other blocks to new files to avoid unnecessary data copy,and pre-reads required data blocks to accelerate the execution.Secondly,to simplify the I/O stack and improve I/O performance,OCKV implements a FTL(Flash Translation Layer)tightly coupled with LSM-tree in the user space,bypasses the file system and directly transfers data with OCSSD,and distributes the file data blocks evenly in all channels of OCSSD when allocating physical space,making full use of the parallelism of multi-channel.Finally,to reduce the write amplification of garbage collection of OCSSD,OCKV separates the physical cold/hot data according to the decreasing heat of files in LSM-tree from the upper layer to the lower layer,and to repair the damage of physical cold/hot separation caused by data block remapping,OCKV migrates the valid flash pages belonging to different layers of LSM-tree to different storage areas in the process of garbage collection and maintains the characteristics of physical cold/hot separation.From the test results,compared with the traditional key value store based on LSM-tree,OCKV can reduce the write amplification by 1.28 times,increase the throughput by 2.47 times to 2.86 times under light write workload,and increase the throughput by 3.1 times to 3.4 times under heavy write workload.With the help of hot/cold data separation according to the features of LSM-tree,OCKV can reduce the data migration caused by garbage collection of OCSSD by 11.8%,improving the efficiency of garbage collection.