BDAaaS模式下面向近似查询的任务调度建模及优化算法  

Task Scheduling Modelling and Optimization Algorithms for Approximate Query Service in Big Data Analytics-as-a-Service

作　　者：叶志 程岩[1] YE Zhi;CHENG Yan(School of Business,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学商学院,上海200237

出　　处：《工业工程与管理》2024年第2期59-69,共11页Industrial Engineering and Management

基　　金：国家自然科学基金资助项目(71271087)。

摘　　要：大数据分析即服务(big data analytics-as-a-service,BDAaaS)是基于云平台提供数据分析服务的新型业务模式,很多企业基于其业务特点愿意牺牲精确度换取更快且费用更低的查询结果。因此,BDAaaS可用数据分割等技术提供近似查询服务。用户对近似查询服务的响应速度要求高,导致服务水平协议(service level agreement,SLA)中对服务响应的约束也更严格,BDAaaS面临更大的违约赔付风险。首先,在服务响应速度要求高、计算资源有限、部分子任务具备顺序依赖关系的前提下,对于近似查询服务中的云任务调度问题,构建了面向收益优化的近似查询任务调度模型。其次,为解决模型中不同子任务的顺序依赖难题,提出了一种基于遗传算法与强化学习相融合的优化算法。该算法利用遗传算法求解带依赖关系的组合优化问题,强化学习在求解过程中动态调整遗传算法的关键参数,在SLA赔付风险下保证调度效率与质量。最后,以仿真实验模拟不同场景来验证所提算法的求解效果与性能。结果表明,该算法可在计算资源有限的条件下,获得SLA违约率更低且收益更高的任务调度方案,从而确保用户满意度和BDAaaS运营商的收益水平。Big data analytics-as-a-service(BDAaaS)is a new business model for providing data analytics services based on a cloud platform.Many companies are willing to sacrifice accuracy for faster and less expensive query results due to their business characteristics.BDAaaS can provide approximate query services by processing large data sets into small representative data samples using techniques such as data splitting.The high response time required by users for approximate query service leads to stricter constraints on service response in service level agreement(SLA),which exposes BDAaaS to a greater risk of default payments.Under the premise of high service response speed requirements,limited computing resources,and some tasks having sequential dependencies,a revenue-optimized approximate query task scheduling model was first constructed for the cloud task scheduling problem.Second,an optimization algorithm based on genetic algorithm and reinforcement learning was proposed to solve the sequential dependency problem of different tasks in the model.Genetic algorithm solved combinatorial optimization problems with dependencies.Reinforcement learning dynamically adjusted the key parameters of the genetic algorithm during the solution process,which was used to ensure the efficiency and quality of scheduling under strict SLA.Finally,different scenarios were simulated to verify the effectiveness and performance of the proposed algorithm.The results show that the algorithm can achieve task scheduling solutions with lower SLA violation rates and higher revenue with limited computing resources,thus ensuring user satisfaction and revenue levels for BDAaaS.

关 键 词：大数据分析即服务 近似查询 任务调度 遗传算法 强化学习 

分 类 号：C931[经济管理—管理学]