【作者】 王睿；

【导师】 戚正伟；

【作者基本信息】 上海交通大学 ， 软件工程， 2011， 硕士

【摘要】 目前的数据中心大多应用虚拟化技术使多个企业级应用共享同一个管理平台,通过这样的方法来增加资源和空间的利用率。但是对于那些对资源的需求不仅随资源不同而改变,而且随时间不同而改变的应用,虚拟化技术在满足它们的应用级服务质量(QoS)目标方面做的还远远不足。另外,当共享服务器出现过载或者为应用设置不同的资源请求优先级的情况时,资源分配将会面临更多的问题。本文主要介绍一个基于反馈的动态资源分配系统,这个系统正是针对这类复杂应用,通过实时动态的调整分配给每个虚拟机的资源,同时生成分派给每个虚拟服务器的负载以达到应用的服务级性能目标(application-level SLOs),这是本系统的主要创新点。该控制系统的参数主要基于该系统的两大部分随系统运行实时调节的,第一个部分是一个在线模型估计器,主要是通过系统识别和递推最小二乘法(RLS)来动态的建立应用的性能和分配给每个虚拟服务器的资源之间的关系;第二个部分是一个优化器,主要是通过一个二次的误差函数以及第一部分建立的性能和资源之间的关系模型来决定分配给每个虚拟服务器的一个最优资源量以达到我们的性能目标。为了评估该控制系统,我们建立了一个测试平台,该平台应用Xen虚拟机监控器(VMM)来管理若干个虚拟机,并且使用Apache服务器以及它的负载生成器,使测试环境尽量贴近实际。通过实验结果能够发现我们的控制器能够检测到虚拟机服务器的资源请求随时间的变化以及分配给其相应的资源以达到应用的服务质量目标。本文的创新部分主要包括以下几点:1.动态资源与性能关系模型:系统在运行过程中动态建立虚拟机资源与性能之间的关系模型,通过细粒度的线性关系模型来估计非线性模型,以使模型精确化。2.动态资源分配系统:系统利用历史数据来预测出下一时段分配给每个虚拟机的一个最优资源,以使其性能达到预设目标。3.集成化虚拟数据中心测试平台:通过该平台测试得出我们的系统不仅能够很好的达到预设的应用服务质量目标,在稳定性和有效性方面都远远优于Xen本身的静态资源分配策略。该平台集成了Xen虚拟机和Apache服务器等关键技术,能够广泛应用在相关的虚拟机性能分析和系统建模过程中。论文的结构如下:首先对基于虚拟化技术的性能保证及相关方面的国内外研究现状做简单介绍;接着对我们系统用到的一些方法和技术,主要包括:虚拟化技术、系统识别和建模过程和最优控制理论等核心算法做详细分析介绍;然后提出整个控制系统的总体架构设计及各个主要模块的功能及实现方法;最后详细描述和分析该控制系统在真实虚拟数据中心平台测试后的实验环境。全文详细介绍了这个基于反馈控制理论和虚拟化技术的虚拟服务器动态资源分配系统的设计与实现过程,该系统对于以后的虚拟化技术的性能优化及系统建模方面的研究,具有一定的价值。更多还原

【Abstract】 Today’s data center often employs virtualization to allow multiple enterprise ap-plications to share a common hosting platform in order to improve resource and spaceutilization. However, current virtualization technologies are inadequate in achievingapplication-level quality of service (QoS) goals with time-varying demands for mul-tiple resources. In addition, it could be a greater challenge when the shared serverbecomes overloaded or enforces priority of the common resource among applicationsin such an environment.In this paper, we propose a feedback-based dynamical resource allocation systemwhich is applied for such complex application can dynamically adjust the resourceshares to multiple virtual machine nodes and workload generation schemes in order toachieve application-level SLOs on shared virtualized infrastructure, which is the mostimportant innovation of the system. The control system parameters are automaticallytuned at the system’s runtime based on two control part. The first part is an onlinemodel estimator using a recursive least-squares (RLS) method to capture the complexrelationship between application performance and resource allocations. The second isan optimal controller using a quadratic cost function to allocate the optimal amountof resources to ensure application-level QoS. To evaluate our controller design, webuilt a testbed hosting several virtual machines which employed Xen Virtual MachineMonitor (VMM), and using Apache server along with its workload-generate tool. Ourexperiment results indicate that our controller is able to detect and adapt to resourcerequirement that changes over time and allocate virtualized resources accordingly toachieve application-level QoS.The innovations of the paper are as follows:1. Dynamic relationship model of resources and performance: The system dynam- ically builds the relationship model between virtual machine resources and per-formance during operation. The system uses fine-grained linear relationshipmodel to approximate non-linear model, which can make the model much accu-rate.2. Dynamic resource allocation system: The system uses historical data to predictan optimal resource allocated to each virtual machine next time to meet the pre-set performance goal.3. Integrated virtual data center test platform: From the platform’s testing, Wecan find that our system no only can meet the pre-set QoS target, but also itsstability and efficiency are far superior than the static resource allocation of Xenown strategy. The test platform integrates Xen virtual machine, Apache serverand other key technologies, which make it can be widely applied in the relevantvirtual machine performance analysis and system modeling process.The paper is structured as follows: first giving a brief introduction of the researchstatus of the performance of virtualization technology ensuring and related areas. Thentaking a description in detail of some of our methods and techniques the system used,such as virtualization technology, system identification and modeling process, opti-mal control theory and other core algorithms. Next presents the general framework ofthe whole control system design and various functions of main modules and the imple-mentation methods. At last we describe and analysis the experiment results of the con-trol system in a real test virtual data center platform experimental environment. Thispaper fully describe the design and implementation of a dynamic resource allocationsystem of virtual servers which based on feedback control theory and virtualizationtechnology. It has a certain value for future research on performance optimization ofvirtualization technology and system modeling.更多还原