LATuner: An LLM-Enhanced Database Tuning System Based on Adaptive Surrogate Model

Database Management Systems (DBMSs) offer a plethora of configurable parameters—termed “knobs”—that control the system behavior. Identifying the optimal configuration for these knobs, i.e., Knob Tuning (KT) is acknowledged as a critical way to enhance the DBMS performance. However, the increasing number of adjustable knobs and the complexity inherent in KTs have rendered manual tuning an antiquated and impractical approach. Recently, automatic KTs based on Machine Learning (ML) techniques have demonstrated significant potentials. Despite the advancements, they are also hindered by notable drawbacks such as the lack of domain knowledge and low tuning efficiency. Meanwhile, Large Language Model (LLM), which is pre-trained on diverse corpora including web content, database manuals, could offer a novel, training-free approach to significantly mitigate the aforementioned issues. In light of this, we propose an LLM-enhanced dAtabase Tuner, called LATuner. Firstly, since KT often suffers from the cold-start problem, we harness the extensive domain knowledge of LLMs to identify critical knobs and to warm start the tuning process, thus obtaining high-quality training samples. Secondly, as KT requires multiple rounds of sampling during the training process, we leverage LLMs to guide the sampling procedure, accelerating the convergence of the model training. Finally, to balance the tuning cost and efficiency between LLM-based KT and traditional ML-based KT, we design an adaptive surrogate strategy based on multi-armed bandit, achieving cost-effective tuning performance. Extensive experiments performed on well-established benchmarks have proven the efficacy and superiority of our proposal.