XTuning: Expert Database Tuning System Based on Reinforcement Learning

Abstract

Database performance optimization has become a hot issue in recent years. Some works deeply reconstruct the database to achieve specified goals like throughput or latency. The others focus on the database’s configuration knobs with reinforcement learning (RL) to improve the performance without any empirical knowledge. But the exhaustive offline training process costs plenty of time and resources due to the large inefficient configuration knobs combinations with trial-and-error methods. The most time-consuming part of the process is not the RL network training, but the database performance evaluation for acquiring the reward values of target performance like throughput or latency. So we propose an expert database tuning system (XTuning) which contains a correlation knowledge model to remove unnecessary training costs and a multi-instance mechanism (MIM) to support fine-grained tuning for diverse workloads. The models define the importance and correlations among these configuration knobs for the user’s specified target. Then we implement the models as Progressive Expert Knowledge Tuning (PEKT) algorithm with an abstracted architectural optimization integrated into XTuning. Experiments show that XTuning can effectively reduce the training time and achieves extra performance promotion compared with the state-of-the-art tuning methods.

This work is supported by the National Key Research and Development Program of China (No. 2019YFE0198600), National Natural Science Foundation of China (No. 61972402, 61972275, and 61732014).