Improving Write Performance for LSM-tree-based Key-Value Stores with NV-Cache

Abstract

Log Structured Merge tree (LSM-tree) is widely used in modern key-value stores due to its superior write performance. However, current LSM-tree-based key-value stores suffer from two performance issues: (1) write stall caused by speed gap between fast DRAM and slow storage devices; (2) write amplification incurred by existing LSM-tree compaction. This paper proposes NV-Cache, a write-optimized non-volatile cache for LSM-tree-based key-value stores. We design a split LSM-tree over a hybrid storage hierarchy consisting of fast Non-Volatile Memory (NVM) and slow solid state drive (SSD). It promotes frequently-accessed, high LSM-tree levels onto fast NVM and offloads cold, large LSM-tree levels into the slow storage device. Further, we propose a range of techniques, including unsorted LSM-tree node layout, lightweight MemTable flush, write-optimal list compaction, and LRU-based NV-Cache eviction, to extensively optimize the write path. NV-Cache closes the performance gap between fast DRAM and slow storage devices to prevent the write stall and reduce write amplification. We implement NV-Cache based on Google LevelDB and conduct a series of experiments using db_bench and YCSB workloads. Experiment results suggest that NV-Cache effectively prevents write stall, achieves 2.3× write reduction, and performs 54% better than LeveIDB., NoveLSM., and MatrixKV.