基于强化学习的SBS云应用自适应性能优化方法  

A Reinforcement Learning-Based Self-Adaptation Performance Optimization Approach for SBS Cloud Application

作　　者：闫永明[1] 张斌[1] 郭军[1] 孟煜[1] 

机构地区：[1]东北大学计算机科学与工程学院,沈阳110819

出　　处：《计算机学报》2017年第2期464-480,共17页Chinese Journal of Computers

基　　金：国家自然科学基金(61572117;61300019;61370155);省科技项目攻关项目(2015302002);中央高校东北大学基本科研专项基金(N140406002;N150404008)资助~~

摘　　要：自适应的调整云应用所占用的资源是一种有效的保障云应用性能的方法,但传统的决策方法面向基于服务的系统(Service-Based System,SBS)时会存在一些问题,例如基于应用系统性能模型的决策方法不能很好适应云环境下SBS的动态变化,基于智能优化算法的决策方法效率较低.该文提出了一种基于强化学习的SBS云应用自适应性能优化方法.在该方法中,该文建立了自适应基本要素之间相互关系的特征描述框架,利用高层次的系统行为指标(如响应时间、用户并发量、资源量等)来描述系统性能的优化目标等.为了应对云环境以及SBS的动态变化,该文的方法采用了无模型(model-free)的在线学习算法,当用户并发量发生变化导致系统的预期行为发生偏差时,该方法通过不断重复"执行-积累-学习-决策"的过程,可以不断的积累经验数据并优化决策结果.为了保证自适应优化的高效性,该文提出了一种引导算子,可以有效的缩小候选自适应动作的范围,提高算法的学习效率.该文实现了以一个SBS为例的原型框架,使用该框架的实验结果证明了该文提出方法的有效性.Adaptive adjustment of cloud applications resources is an effective way to guarantee the performance of cloud applications. There are some problems of the traditional decision method for SBS (Service-based System), such as the decision method based on application system performance model cannot adapt to the dynamic change of SBS in cloud environment, the efficiency of the decision making method based on intelligent optimization algorithm is not high. To solve the disadvantages of the existing methods, a self-adaptation performance optimization approach for SBS cloud application based on reinforcement learning is proposed by this paper. In this approach, the feature description framework of the relationship of the basic key adaptation elements is established, and the system's optimization target for system performance and key elements of learning algorithm are described as high-level system behavior indicators (such as response time, user concurrency, resource quantity, etc.). To cope with dynamic change of the cloud environment and SBS, this paper adopts a model-free online learning algorithm. When the cloud application deviates from the desired behavior caused by the changes in the number of user concurrency, the approach will repeat the 'execution-accumulation-learning-decision' process, which can accumulate the corresponding empirical data in the process of dynamic decision, and optimize the decision-making results continuously through empirical data. To ensure the high efficiency of the adaptation optimization, this paper puts forward a shaping operator which can effectively narrow the scope of the candidate adaptation actions and improve the learning efficiency of the algorithm. Experimental results using a prototype framework in the context of a SBS application demonstrate the effectiveness of this approach. © 2017, Science Press. All right reserved.

关 键 词：自适应 强化学习 资源调整 云应用 基于服务的系统 云计算 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]