面向能效的云计算虚拟化资源提供方法研究

发布时间：2018-06-08 05:02

本文选题：能效 + 虚拟化　；参考：《国防科学技术大学》2015年博士论文

【摘要】：随着云计算产业的迅猛发展,大规模数据中心的能耗问题变得日益严峻。工业和学术界都开始努力尝试使用各种技术以提高数据中心的能效。其中,最重要同时也是云计算采用的最关键技术之一就是资源的虚拟化。相对于非虚拟化系统,虚拟化技术带来的弹性能显著提高系统资源利用率,达到节能减排的目的。然而,对共享资源的竞争也使得虚拟机内的应用程序性能受到威胁,这直接影响到云服务的质量,进而影响服务提供商的收益和最终用户体验。如何在这样的虚拟化平台上实现性能和能耗的最佳折衷,是当前业界最关注的问题之一。本文围绕云计算环境中虚拟化资源的提供方法展开研究,旨在通过高效的动态资源提供策略持续优化应用程序资源配置,有效应对应用程序负载波动,保障应用程序性能,提高云计算数据中心资源利用率,同时降低数据中心能耗,最终达到提高系统能效的目的。本文的主要研究成果包括:1.提出了一种基于支持向量机的负载预测方法虚拟化是提高数据中心能效的主要技术手段。而云环境中很多应用程序通常具有高度变化的负载,导致复杂动态的资源使用模式。资源提供策略必须根据应用程序的实际需求实时动态地调整虚拟机资源的供给量,才能在满足QoS约束的前提下尽量减少资源占用量。因为资源的动态调整总是存在延迟,所以,实现以上目标的一个必要前提就是要提前获知虚拟机的未来资源需求。本文提出了一种负载预测方法KSwSVR。该方法将虚拟机的负载预测当作一个时间序列预测问题来解决,并采用了专门针对小样本统计问题而建立的统计学习技术。基于结构风险最小化准则的支持向量机具有很强的泛化能力,因而能有效适应复杂动态的云计算环境。受局部性原理启发,本文对标准的支持向量回归算法进行了改进,区别对待训练数据,根据其重要性进行加权优化,并与卡尔曼平滑技术进行了整合。实验结果表明,KSwSVR能有效预测各种资源类型的负载对象,并且在预测精度、稳定性、算法开销等方面均优于常用经典算法。2.提出了一种基于负载预测的资源自动扩展方法云计算最重要的特征是弹性,应用程序可以根据需求动态地获取或释放资源。从管理系统的角度,就是要利用虚拟化提供的资源动态扩展能力,使得应用程序的资源配置能尽量实时精确地匹配应用程序的需求变化。同时,云计算系统的规模要求扩展操作必须是自动化的,整个过程要减少甚至完全消除人的干预。精确的负载预测结果可以用于指导资源扩展方案的制定,但是由于预测误差的存在,直接将预测值作为资源供给量会导致应用程序性能不稳定。本文以保证应用程序性能、最小化资源占用量为目标,提出了一种新的资源自动化扩展方法G2LC。该方法综合实验经验,并通过接收时间窗口内的QoS反馈,从全局增益和局部误差补偿两个层面对预测结果进行了修正。在满足SLA约束的同时,有效减少了瞬时负载带来的不必要扩展动作,完全避免了SLA违约事件的集中出现。实验结果表明,G2LC可以通过调整参数有效地保证用户指定的任意性能水平。相对于虚拟机资源固定不变的定值分配方法,G2LC能显著地节省资源开销,且这一优势随着用户性能要求的提高变得越发明显。3.提出了一种基于服务质量差异化的资源提供方法性能管控和节能减排是现代数据中心的两个主要研究课题,但两者又往往是相矛盾的。云服务提供商运用虚拟化技术将各种应用程序整合到更少的物理主机上,并将闲置的主机转换到低功耗模式以降低能耗。这种情况下,应用程序的性能严重依赖于对虚拟机容量的合理管理。因为应用程序的多样性、各种资源使用模式的差异、底层硬件共享、程序之间的性能相关和性能干扰等因素的存在,导致对虚拟机性能的管控非常困难。找到能耗和性能之间的平衡点至关重要,这被归结为能效问题。本文最后以提高系统的能效为目标,基于对负载预测技术和资源扩展技术的研究成果,提出了一种新的资源提供方法CoST。它基于服务质量差异化策略,利用了不同类型的应用程序对性能和成本的敏感度不同这一事实。性能敏感型应用程序追求稳定的QoS,而性能耐受型应用程序更关注完成所有任务所需要付出的总成本。CoST将不同类型的应用程序部署在同一台主机上,基于负载预测和QoS反馈对虚拟机的资源进行垂直扩展。实验表明CoST不仅有效地保证了性能敏感型应用程序的QoS,支持云用户在应用程序运行过程中在线更改性能指标,还可能提高性能耐受型应用程序的总体处理速度。最重要的是,CoST维持主机始终运行在最高能效状态,显著地提高了系统总体能效。
[Abstract]:With the rapid development of the cloud computing industry, the problem of energy consumption in large-scale data centers has become increasingly severe. Industry and academia have begun to try to use various technologies to improve the energy efficiency of data centers. One of the most important technologies used in the cloud computing is the virtualization of the resources, compared with non virtualized systems. The flexibility brought by the virtualization technology can significantly improve the system resource utilization and achieve the purpose of energy saving and emission reduction. However, the competition of shared resources also threatens the performance of the application in the virtual machine, which directly affects the quality of the cloud services, and then affects the benefits of the service providers and the end user experience. The best tradeoff between performance and energy consumption on the pseudo platform is one of the most important issues in the current industry. This paper focuses on the method of providing virtualized resources in the cloud computing environment, aiming at optimizing the allocation of application resources through efficient dynamic resource provision, effectively responding to the application load fluctuation and ensuring the application process. Sequence performance improves the utilization of cloud computing data center resources and reduces the energy consumption of data center at the same time, and finally achieves the purpose of improving the energy efficiency of the system. The main research results of this paper include: 1. proposed a support vector machine based load forecasting method virtualization is the main technical means to improve the energy efficiency of the data center. The program usually has a highly variable load that leads to complex dynamic resource usage patterns. The resource provision strategy must dynamically adjust the supply of virtual machine resources according to the actual requirements of the application in order to reduce the resource occupancy as much as possible under the premise of satisfying the QoS constraints. Therefore, a necessary prerequisite for achieving the above objectives is to know the future resource requirements of the virtual machine. This paper proposes a load forecasting method KSwSVR. which uses the load prediction of the virtual machine as a time series prediction problem, and uses a statistical learning technique specially designed for small sample statistical problems. The support vector machine based on structural risk minimization criterion has strong generalization ability and can adapt to the complex dynamic cloud computing environment effectively. Inspired by the locality principle, this paper improves the standard support vector regression algorithm, treats the training data differently, and optimizes its importance according to its importance, and is flat with Calman. The experimental results show that KSwSVR can effectively predict the load objects of various resource types, and is superior to the common classical algorithms in prediction accuracy, stability, algorithm overhead and so on..2. proposed a resource automatic extension method based on load forecasting. Resources are dynamically acquired or released according to the requirements. From the management system point of view, the resource dynamic extension capability provided by virtualization enables the resource configuration of the application to match the needs changes of the application as accurately as possible. At the same time, the scale of the cloud computing system requires that the extended operation must be automated and complete. The accuracy of the load prediction results can be used to guide the formulation of the resource expansion scheme, but due to the existence of the prediction error, the direct prediction value as a resource supply will lead to the instability of the application performance. This paper aims to ensure the application sequence performance and minimize the resource occupancy. A new method of resource automation extension G2LC. is presented, which combines the experimental experience with the QoS feedback in the time window and corrections the prediction results from two layers, the global gain and the local error compensation, which can effectively reduce the unnecessary expansion action brought by the instantaneous load and completely avoid the SLA constraints. The concentration of SLA default events. The experimental results show that G2LC can effectively guarantee the user's arbitrary performance level by adjusting the parameters. Compared with the fixed value allocation method of virtual machine resources, G2LC can significantly save the resource cost, and this advantage becomes more and more obvious with the improvement of the user's performance requirements. The performance management control and energy saving and emission reduction based on the quality of service quality differentiation are two main research topics in the modern data center, but they are often contradictory. In this case, the performance of the application depends heavily on the reasonable management of the capacity of the virtual machine. Because of the diversity of the application, the differences in various resource usage patterns, the underlying hardware sharing, the performance related and performance interference among the programs, the control of the performance of the virtual machine is very poor. Finding the equilibrium point between energy consumption and performance is essential, which is attributed to energy efficiency. Finally, this paper aims at improving the energy efficiency of the system. Based on the research results of load forecasting and resource expansion, a new method of resource provision, CoST., is proposed based on the quality of service differentiation strategy and the use of different types. The fact that applications are sensitive to performance and cost is different. Performance sensitive applications pursue stable QoS, while performance tolerance applications pay more attention to the total cost that is required to complete all tasks.CoST deploys different types of applications on the same host, based on load forecasting and QoS feedback to virtual machines The experiment shows that the CoST not only effectively guarantees the QoS of the performance sensitive application, but also supports the cloud users to change performance metrics online in the application process, and can also improve the overall processing speed of the performance tolerant application. Most importantly, the CoST mainframe is always running in the highest energy efficiency state, The overall energy efficiency of the system has been significantly improved.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP393.09

【相似文献】