Research On LSM-Tree Key-Value Stores Based On Address Remapping

Solid State Drive(SSD)and Log-Structured Merge Tree(LSM-tree)based key-value stores have been widely used due to their high performance and scalability.The problem of double write amplification in LSM-tree key-value stores caused by compaction and garbage collection severely degrades system performance and accelerates SSD wear.Reusing data in compaction by address remapping is an effective way to reduce write amplification.However,existing works do not propose how to detect the reusable data in compaction effectively and how to improve the reuse rate of these data.To reduce the write amplification of LSM-tree key-value stores,we design and implement a key-value store system based on address remapping called Remap KV.Remap KV efficiently detects the reusable data according to the characteristics of key-value pairs in compaction.To improve the reuse rate of these data and avoid duplicate data writes,Remap KV calls address remapping command in alignment.Based on Open-Channel SSD,Remap KV transfers the reusable information of data from key-value stores to FTL on the host side via address remapping command.Remap KV remaps the logical address of repeatedly written data to the physical address of existing data in FTL.To support the manyto-one relationship between logical address and physical address caused by address remapping,address management policy and garbage collection policy are proposed.Experimental results show that compared with Rocks DB,for random updates the throughput of Remap KV is improved by 20%,the average latency is reduced by 17%,the total write amplification is reduced by 42% and the times of erasing are reduced by 31%.For write-intensive workloads,the throughput of Remap KV is increased by 32%,the average latency is decreased by 24%,the 99.9% tail latency is decreased by 2.4 times,and the write amplification is decreased by 22%.For read-only workloads,the throughput and the average latency of Remap KV do not deteriorate,indicating that Remap KV do not sacrifice read performance while reducing write amplification.