Speculative Multi-Level Access in LSM Tree-Based KV Store

Abstract

The log-structured merge (LSM) tree has been widely used due to its small DRAM footprint, easy range query support, and write-friendly characteristics. However, it has an inherent problem of read amplification; a single key lookup operation can be accompanied by multiple disk read operations which prolong the latency of lookup operation and fail to exploit the parallelism of SSDs. This paper proposes a speculative multi-level access scheme, which accesses multiple levels simultaneously and speculatively. By accessing multiple levels together, it can exploit multi-I/O-handling capability of the SSD and also overlap the execution of CPU with I/O device, thereby improving the performance of lookup operations. Our evaluation results demonstrate that the proposed scheme improves the throughput by up to 85% and reduces the average latency by up to 45% with realistic key-value workloads.