Abstract:Key-value (KV) storage systems based on log-structured merge-tree (LSM-tree) architectures are organized hierarchically, where an increasing number of levels exacerbates read and write amplification. To address this issue, this study proposes Split-LSM-Tree, an optimized KV storage engine that dynamically partitions a single LSM-tree into a forest of sub-trees, thus effectively reducing the sub-tree depth. Specifically, a load-aware adaptive mechanism is designed to precisely trigger splitting. Furthermore, a non-blocking split algorithm combined with I/O optimizations ensures uninterrupted front-end read and write operations during splitting. Experimental results show that Split-LSM-Tree reduces write amplification by 25.7% and 18.1% compared to LevelDB and RocksDB, respectively, while increasing write throughput by 2.9× and 2.4×. Under typical YCSB workloads, its overall throughput is improved by up to 30% compared with the KV-separated engine WiscKey.