Database Parameters Tuning via Bayesian Optimization with Domain Knowledge

Database Parameters Tuning via Bayesian Optimization with Domain Knowledge

연구 분야: Databases

학회: International Conference on Web Information Systems and Applications

Bayesian optimization has gained widespread adoption in database knob tuning due to its theoretical advantages in balancing exploration and exploitation. Yet, a significant drawback of existing Bayesian optimization-based approaches is typically their failure to incorporate domain knowledge related to databases when searching for the optimal configuration. This limitation often leads to the recommendation of low-utility configurations that violate domain knowledge, thereby affecting its tuning efficiency. To address this issue, we propose DKTune, which seamlessly integrates Bayesian optimization with domain-specific database knowledge. DKTune leverages the inherent dominant relationships between database knobs to enhance the surrogate model used in Bayesian optimization. Additionally, it considers constraint relationships between knobs, competitive interactions among knobs, and the dynamic characteristic of knobs to assist the acquisition function in evaluating the utility of each configuration. We evaluated DKTune on two popular open-source database systems, and the experimental results demonstrate that DKTune significantly improves the efficiency of database knob tuning and the final tuning results.