Reproduce “MCTuner: Spatial Decomposition-Enhanced Database Tuning via LLM-Guided Exploration” on OpenResearch

About this paper

Zihan Yan, Rui Xi, Mengshu Hou

MCTuner, developed by researchers at the University of Electronic Science and Technology of China, is a database tuning framework that uses multi-modal large language models (LLMs) and spatial decomposition to efficiently optimize database performance. It significantly reduces exploration in ineffective configuration regions, achieving up to 19.2% performance improvement over state-of-the-art methods and faster configuration discovery across diverse workloads.

Key insights

• Combining LLM-guided knowledge for initial knob selection with dynamic spatial decomposition effectively prunes the configuration space, drastically reducing exploration of irrelevant or suboptimal regions.
• A Mixture-of-Experts (MoE) architecture enhances knob selection interpretability and accuracy by leveraging multi-source domain knowledge and distributed LLM evaluation.
• Integrating MCTS with localized Bayesian Optimization within a recursively partitioned space allows for robust and efficient discovery of high-performance configurations, even with limited runtime samples.

Summary by alphaXiv.