基于合作关系的多智能体数据库参数调优模型

Multi-agent Database Parameter Optimization Model Based on Cooperative Relationships

高维数据库参数空间中的参数调优是提高数据库性能的难点,现有方法更多关注于如何识别重要参数,在如何有效提高可调参数数量的问题上仍存在不足。针对上述问题,基于低维度映射技术和多智能体(Multi-Agent)强化学习技术,提出基于合作关系的Multi-Agent数据库参数调优(Cooperative Multi-Agent Database Parameter Tuning,CMADPT)模型,CMADPT将数据库参数进行分类调优,极大增加了可调参数的数量;提出基于低维度映射的降维模型(Low Dimensional Mapping Model,LDMM),通过低维的合成参数调优高维的数据库参数。实验结果表明,CMADPT模型有效地扩大了可调参数的数量,比主流方法平均提升1.117%的数据库性能。此外,CMADPT每300次迭代训练平均节省1.32 h,极大地提升了算法的时间性能。

Parameter tuning in high-dimensional database parameter space is a difficult problem to improve database performance.Existing methods have drawbacks on how to effectively expand the number of tunable parameters,and these methods focus more on how to identify important parameters.To address above problems,a Multi-Agent-based database parameter tuning model called CMADPT(Cooperative Multi-Agent Database Parameter Tuning)is proposed based on low-dimensional mapping and Multi-Agent reinforcement learning techniques,which classifies database parameters for tuning and greatly increase the number of tunable parameters.A Low Dimensional Mapping Model(LDMM)is proposed,which tunes high-dimensional database parameters through low-dimensional synthetic parameters.Experimental results show that the CMADPT model can effectively expand the number of tunable parameters and improve the database performance by 1.117%on average when compared with the state-of-the-art methods.In addition,CMADPT can save 1.32 h per 300 iterative training on average,which can greatly improve the runtime performance of the proposed model.