GPTuner: An LLM-Based Database Tuning System

• DOI: 10.1145/3733620.3733641
• Publication Date: 2025-04-29T16:20:07Z
• Authors: Jiale Lao; Yibo Wang; Yufei Li; Jianping Wang; Yunjia Zhang; Zhiyuan Cheng; Wanghu Chen; Mingjie Tang; Jianguo Wang
• Keywords: Bayesian Optimization; Design space exploration

Selecting appropriate values for the configurable knobs of Database Management Systems (DBMS) is essential to improve performance. But because complexity this task has surpassed abilities even best human experts, database community turns machine learning (ML)- based automatic tuning systems. However, these systems still incur significant costs or only yield suboptimal performance, attributable their overly high reliance on black-box optimization and lack integration with domain knowledge, such as DBMS manuals forum discussions. Hence, we propose GPTuner, a manual-reading system that extensively leverages knowledge automatically optimize search space enhance runtime feedback-based process. Firstly, develop Large Language Model (LLM)-based pipeline collect refine heterogeneous prompt ensemble algorithm unify structured view refined knowledge. Secondly, using (1) design workload-aware, trainingfree knob selection strategy, (2) technique considering value range each knob, (3) Coarse-to-Fine Bayesian Optimization Framework explore optimized space. Finally, evaluate GPTuner under different benchmarks (TPC-C TPC-H), metrics (throughput latency) (PostgreSQL MySQL). Compared state-of-the-art methods, identifies better configurations in 16x less time average. Moreover, achieves up 30% performance improvement over best-performing alternative.