Mitigating Resource Usage Dependency in Sorting-based KV Stores on Hybrid Storage Devices via Operation Decoupling

Explore a 15-minute conference presentation from USENIX ATC '25 that addresses critical performance issues in LSM-tree-based key-value stores on hybrid storage devices. Learn how sorting-based operations like flush and compaction create operation coupling problems, including intertwined resource consumption, interdependencies, and contention that lead to resource fragmentation and increased write stalls. Discover DecouKV, an innovative solution that decouples data sorting operations into CPU-intensive index merge tasks and I/O-intensive data append and flush tasks by separating indexes from data files. Understand how DecouKV manages indexes with a mergeable skip list-based structure and data with append-only files, while implementing an elastic scheme for tuning level capacity and a parameterized queue-based task scheduling strategy. Examine experimental results showing DecouKV's significant improvements over RocksDB and state-of-the-art systems like MatrixKV, PrismDB, SplitDB, and ADOC, including 25.4%-32.3% better CPU utilization, 2.3-4.9× throughput increases, and 74.3%-91.4% tail latency reductions under write-intensive workloads, plus modest 1.2-2.3× throughput improvements for read-intensive scenarios.