数据库参数调优是提高数据库性能的重要任务之一.数据库参数可按照作用域范围和功能进行分类,探究某一类参数之间的相互影响以及不同类别之间的相互影响是非常重要的,但是现有方法尚未考虑这一方面.提出面向数据库参数调优的协作型多智能体模型DBT-MADDPG(DataBase Tuning-Multi-Agent Deep Deterministic Policy Gradient),设计单智能体预训练模型SA(Single Agent)、多智能体联合训练模型JAM(Joint Action Model)以及基于概率选择的联合训练模型JAPM(Joint Action Probability Model),分阶段对数据库参数进行调优.实验结果表明,DBT-MAD-DPG模型能够分功能、分参数级别对数据库参数进行调优,在SA模型训练阶段便能到达主流算法的性能,且比主流算法快17.85%获得较佳配置.
A Collaborative Multi-Agent Model for Database Parameter Tuning
Database parameter tuning is one of the crucial tasks in improving the performance of database systems.Database parameters can be classified based on their scopes and functionalities.It plays an essential role in investigating the mutual influence of parameters within a specific category or between different categories.But,the existing methods do not take into consideration this aspect.A collaborative multi-agent model called DBT-MADDPG(DataBase Tuning-Multi-Agent Deep Deterministic Policy Gradient)is proposed for database parameter tuning.A single-agent pre-training model called SA(Single Agent),a multi-agent joint training model called JAM(Joint Action Model),and a joint training model based on probabilistic selection called JAPM(Joint Action Probability Model)are designed for tuning the database parame-ters at different stages.The experimental results show that the DBT-MADDPG model is capable of tuning the database pa-rameters at different functional and parameter levels,and can reach the performance of mainstream algorithms in the train-ing stage of the SA model,and is 17.85%faster than the state-of-the-art algorithms to obtain the optimal configuration.
multi-agentparameter tuningdatabase systemself-learningjoint training model
万方数据
维普
