CmptQuiet: Optimizing Tail Latency and Throughput in LSM-Tree Key-Value Stores via Contention-Aware Compaction

Abstract

Log-Structured Merge-Tree (LSM-tree) based Key-Value stores are foundational to data center storage but suffer from high tail latency due to a fundamental read-write trade-off and disruptive compaction. This paper introduces CmptQuiet, a system that breaks this trade-off to deliver low tail latency without sacrificing throughput. CmptQuiet employs an auxiliary index table and an adaptive resource allocator to mitigate compaction’s impact. Micro-benchmark results across various Read:Write ratios and access distributions show that CmptQuiet reduces P99 tail latency by up to 91.5% compared to auto-tuned RocksDB, while further improving the system throughput. In macro-benchmarks using real-world YCSB workloads, CmptQuiet outperforms state-of-the-art systems including SILK, CruiseDB, and ADOC. It achieves up to 40.9% lower P99 tail latency than CruiseDB. Moreover, CmptQuiet maintains a stable throughput advantage over all three baselines, with a maximum average improvement reaching up to 22.3%.