基于异步Dueling DQN和计划时间预测网络的连接优化器  

A Join Optimizer Based on Asynchronous Dueling DQN and Plan Latency Prediction Network

作　　者：高瑞玮 乔少杰 韩楠[2] 闵圣捷 李贺[4] 覃晓 张桃 元昌安 GAO Rui-wei;QIAO Shao-jie;HAN Nan;MIN Sheng-jie;LI He;QIN Xiao;ZHANG Tao;YUAN Chang-an(School of Software Engineering,Chengdu University of Information Technology,Chengdu,Sichuan 610225,China;School of Management,Chengdu University of Information Technology,Chengdu,Sichuan 610103,China;CEC Zhiyuan Data Technology Co.,Ltd.,Beijing 610225,China;School of Computer Science and Technology,Xidian University,Xi’an,Shaanxi 710126,China;Nanning Normal University,Nanning,Guangxi 530001,China;Yibin University,Yibin,Sichuan 644000,China;Guangxi Academy of Sciences,Nanning,Guangxi 530007,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]成都信息工程大学管理学院,四川成都610103 [3]中电智元数据科技有限公司,北京100081 [4]西安电子科技大学计算机科学与技术学院,陕西西安710126 [5]南宁师范大学,广西南宁530001 [6]宜宾学院,四川宜宾644000 [7]广西科学院,广西南宁530007

出　　处：《电子学报》2023年第7期1868-1874,共7页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62272066,No.61962006);四川省科技计划资助(No.2021JDJQ0021,No.2022YFG0186);教育部人文社会科学研究规划基金(No.22YJAZH088);成都市“揭榜挂帅”科技项目(No.2022-JB00-00002-GX,No.2021-JB00-00025-GX);成都市技术创新研发项目(No.2021-YF05-02413-GX,No.2021-YF05-02414-GX);中国电子科技集团公司第五十四研究所高校合作课题(No.SKX212010057);成都信息工程大学科技创新能力提升计划(No.KYTD202222)。

摘　　要：连接顺序选择是查询优化领域中极具挑战性的研究方向,对于数据库管理系统获得良好的查询性能至关重要.然而,传统优化方法和现有智能优化方法均存在着不足,如规划时间过长、容易得到质量较差的连接计划、编码未考虑结构特征、依赖基数估计和代价估计使得连接计划无法反映真实的执行时间等.针对上述问题,提出了一种新型基于异步Dueling DQN(Deep Q-network)和计划时间预测网络的连接优化器:ADP-Join(Asynchronous Dueling DQN and Plan Latency Prediction Network for Join Order Selection).ADP-Join集成了一种新的编码方法,能够区分不同结构的连接计划.ADP-Join设计了计划时间预测网络PLN(Plan Latency Prediction Network)来改善现有基于强化学习优化器的奖励机制.再者,提出异步更新机制改进Dueling DQN模型来提升训练性能和减少训练时间.大量的实验结果表明,在TPC-H和JOB真实数据集上ADP-Join的性能优于现有的智能优化器.Join order selection is a challenging research topic in the field of query optimization,and it is very impor⁃tant for database management system to obtain good query performance.However,both traditional optimization methods and existing intelligent optimization methods have disadvantages such as long planning time,easily to obtain poor quality join plans,encoding without considering structural characteristics,making join plans unable to reflect the real execution time due to dependency on cardinality estimation and cost estimation.In order to solve the above problems,a new join optimizer ADP-Join(Asynchronous Dueling DQN and Plan latency prediction network for Join order selection)is proposed.ADP-Join integrates a new encoding method that can distinguish join plans of different structures.ADP-Join designs a plan latency prediction network to improve the reward mechanism of existing reinforcement learning-based optimizers.Further⁃more,the asynchronous update mechanism is proposed to improve the Dueling DQN model to improve the training perfor⁃mance and reduce the training time.Extensive experimental results show that ADP-Join outperforms existing intelligent op⁃timizers on real TPC-H and JOB datasets.

关 键 词：连接顺序选择 查询优化 连接计划 强化学习 异步更新 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]