基于“推荐-学习”的两阶段数据布局策略-《南京师大学报（自然科学版）》

文章信息/Info

Title:: A Two-Stage Data Placement Strategy Based on Recommendation and Learning

作者:: 梁杨¹; 2; 丁长松¹; 2; 胡志刚³; (1.湖南中医药大学信息科学与工程学院,湖南长沙 410208)
(2.湖南省中医药大数据分析实验室,湖南长沙 410208)
(3.中南大学计算机学院,湖南长沙 410083)

Author(s):: Liang Yang¹; 2; Ding Changsong¹; 2; Hu Zhigang³; (1.School of Informatics,Hunan University of Chinese Medicine,Changsha 410208,China)
(2.Big Data Analysis Laboratory of TCM in Hunan Province,Changsha 410208,China)
(3.School of Computer Science and Engineering,Central South University,Changsha 41

摘要:: 针对数据布局不合理导致云边协同集群服务质量下降和运营开销增加等问题,提出一种基于“推荐-学习”的两阶段数据副本管理机制.鉴于数据密集型移动应用的特点,综合考虑了云边环境下的数据访问延迟和放置代价之间的最优权衡.从理论上构建了副本放置的多目标数学模型,将决策问题描述为具有延迟和成本约束的双目标优化问题; 在“推荐”阶段,设计了一个基于移动预测和反馈优化的副本推荐引擎,减少了副本创建的盲目性; 在“学习”阶段,构建了一个基于异步优势行动者-评论家算法(A3C)的强化学习副本放置规则学习模型,改进了副本服务的全局性能指标.实验结果表明,基于“推荐-学习”的两阶段数据布局策略能够有效地减少等待延迟和节约成本开销,为现代云边协同系统的数据管理服务提供行之有效的解决方案,具有重要的理论意义和应用价值.

Abstract:: Aiming at the problem that the unreasonable data layout leads to the decrease of service quality and the increase of operation cost in cloud-edge collaborative cluster,a two-stage data replica management mechanism based on ‘recommendation and learning'method is proposed. In view of the characteristics of data-intensive mobile applications,the optimal trade-off between data access latency and placement cost in the cloud environment is considered. The multi-objective mathematical model of replica placement is established in theory,and the decision problem is described as a bi-objective optimization problem with delay and cost constraints. In the recommendation stage,a replica recommendation engine based on motion prediction and feedback optimization is proposed to reduce the blindness of replica creation. In the learning stage,a reinforcement learning replica placement rule learning model based on Asynchronous Advantage Actor-Critic algorithm(A3C)is constructed to improve the global performance index of replica service. The experimental results show that the two-stage data placement strategy based on ‘recommendation and learning' method can reduce latency time and save cost overhead effectively and efficiently. Meanwhile,our proposed method provides an effective solution for challenges of data managemen in the real-life cloud-edge collaboration systems,which has important theoretical significance and application value.