云数据中心绿色节能需求的虚拟机负载均衡技术研究

发布时间：2018-03-14 12:22

本文选题：节能高效　切入点：负载均衡　出处：《南京大学》2016年博士论文　论文类型：学位论文

【摘要】：虚拟化技术作为云计算中最关键的核心技术,使得商用云数据中心可以向用户提供按需配置的弹性计算资源。然而,在数据中心的部署中,高性能计算往往是评价的最主要标准,缺少对能源消耗和碳排放的关注。随着云数据中心的规模越来越大,其消耗的电能也十分巨大,这不仅带来了较高的运行成本,而且还产生了较高的二氧化碳排放量。导致数据中心耗电量巨大的一个很重要的原因是计算资源的低效利用。大量数据表明,数据中心中大部分服务器设备事实上处于低负载运行状态,即处于满负荷的10%-50%,而这些设备却要消耗满负载运行时能耗的70%。不仅如此,还带来了额外的冷却系统能耗。在倡导绿色节能、降低碳排放从而实现可持续发展的今天,云数据中心应该不仅仅关注用户服务质量,而应该更加重视提高资源利用率对能耗的影响,实现绿色节能高效的可持续发展目标。但是,除非开发并应用先进的节能资源管理解决方案,否则这些数据预计将进一步快速增长。本论文重点围绕以绿色节能为目标的的云数据中心资源管理问题进行研究,在满足用户服务质量的前提下,通过建立能耗模型,对比分析云环境下主机负载数据的新特点,提出一系列的主机负载检测、虚拟机安置等动态虚拟机均衡算法,使用真实云环境下的主机负载数据进行仿真实验,结果表明,本文提出的多个算法组合与基准参考算法相比,都有明显的性能提升。具体而言,本文的主要工作内容和创新点如下：1.针对高能耗问题,对云计算系统的节能技术进行了分类研究。从静态和动态功率管理,硬件和软件层面,操作系统级、虚拟机级和数据中心级等多个类别展开了深入的分析。在研究过程中,重点从系统资源、优化目标、节能技术、负载特性等多个方面进行展开。重点分析了操作系统级、虚拟机以及云数据中心级三方面节能高效的关键技术。相关分类研究对本文进一步开展以绿色节能为目标的动态虚拟机均衡算法研究具有很强的指导意义。2.提出了云计算环境下一整套分布式动态虚拟机均衡的启发式算法。提出利用中值绝对偏差和四分差来改进传统的设定静态阈值,进一步利用局部回归LR和强局部回归LRR的动态自适应算法进行过载主机探测。提出利用最短迁移时间策略、随机选择策略和最大相关性策略等三种迁移选择算法进行虚拟机迁移选择。将云计算环境中的虚拟机安置问题定义为一个一维装箱问题,设计基于功耗的改进降序最佳自适应启发式算法BFD-PA进行虚拟机安置。通过使用真实Planetlab云数据中心工作负载数据进行仿真实验。结果表明,所提出的结合局部回归算法和最短迁移时间策略来实现虚拟机均衡的综合算法显著优于其他动态虚拟机均衡算法,能够大大降低SLA违例情况和并减少虚拟机迁移的数量。3.提出了基于相空间重构(PSR)和数据组合处理进化方法(EA-GMDH)的组合算法,首次将其用于云计算环境下的主机负载预测领域。利用PSR方法,将一维工作负载的时间序列重构成了多维空间中的时间序列,送入EA-GMDH神经网络,通过训练和学习得到最终的预测结果。通过使用真实Google云数据中心工作负载数据进行仿真实验。本文提出的EA-GMDH算法在平均负载预测和精确负载预测两个方面都好于对比的基准算法：Bayes, EMA, LWMA, LMA, AR, ANN和PP等。4.提出了一种全新的实现云环境下主机负载预测的方法,利用Autoencoder自编码网络和Softmax分类器,将回归问题转换成分类问题,再次提升了负载预测的准确度。通过无监督学习方法成功地将历史窗口中的负载数据特征提取出来。基于这些提取的特征,利用分类器,成功地获得了负载数据的预测结果。同样,通过使用Google云数据中心的工作负载数据作为数据源进行仿真。与基准参照算法ANN、Bayes以及EA-GMDH算法相比,提出的新算法可以获得更高的准确度性能,从而更好地为实现云环境下更为精确的动态虚拟机均衡提供服务和支撑,最终达到在满足QoS要求下降低系统整体能耗的目的。
[Abstract]:Virtualization technology as the core technology of the key cloud computing, makes the commercial cloud data center can provide on-demand computing resources allocation flexibility to the user. However, in the data center deployment, high performance computing often is the main evaluation standard, lack of energy consumption and carbon emissions. With the cloud data center the increasingly large scale, its power consumption is also very great, which not only brings higher operating costs, but also produce carbon dioxide emissions higher. Lead to huge data center power consumption is a very important reason is the inefficient use of computational resources. A large amount of data shows that, in a state of low load operation most in fact, the data center server device, which is in full load 10%-50%, and these devices are consumed at full load energy consumption 70%. not only that, but also brought additional cooling system The system energy consumption. In promoting green energy, reduce carbon emissions and achieve sustainable development today, cloud data center should not only pay attention to customer service, but should pay more attention to improving the utilization rate of resources impact on energy consumption, realize sustainable development of green energy saving and high efficiency. However, unless such resource management the development and application of advanced solutions otherwise, the data is expected to further rapid growth. This thesis focuses on the study of the green energy-saving goal of the cloud data center resource management issues, in order to meet the demand of customer service quality, through the establishment of energy consumption model, comparative analysis of the host cloud environment load characteristics data, put forward a series of host load detection. The virtual machine placement virtual machine dynamic balancing algorithm, using real host cloud environment simulation results show that the load data. Compared with proposed algorithm combined with reference datum, have significantly better performance. Specifically, the main contents and innovations of this paper are as follows: 1.. Aiming at the problem of high energy consumption, energy saving calculation technology system classified research on cloud. From the static and dynamic power management, hardware and software level, operating system class, virtual machine level and data center level and other categories in-depth analysis. In the course of the study, focusing on the system resources, the optimization goal, energy-saving technology, many aspects of the load characteristics were analyzed. The operating system level, three key technologies of energy efficient virtual machine and cloud data center level. Classification of equalization algorithm has a strong guiding significance of.2. put forward a set of distributed cloud computing environment for the further dynamic virtual machine to green energy-saving target The heuristic algorithm of dynamic equilibrium. This virtual machine uses the median absolute deviation and four points to improve the traditional static threshold setting, the further use of dynamic adaptive local regression algorithm LR and strong local regression LRR overload detection. Using the shortest host migration time strategy, random selection strategy and the maximum correlation of three kinds of migration strategies selection algorithm for virtual machine migration. Cloud computing environment of virtual machine placement problem is defined as a one-dimensional packing problem, improved design of descending optimal adaptive heuristic algorithm BFD-PA power based on virtual machine placement. Through the use of real Planetlab cloud data center work load data simulation experiments. The results show that the combination of local regression algorithm and the shortest time migration strategy to achieve comprehensive and balanced algorithm of virtual machine is significantly better than that of the other dynamic virtual machine Equalization algorithm, which can greatly reduce the SLA violation and.3. reduced the number of virtual machine migration is proposed based on phase space reconstruction (PSR) and the combination of data processing methods (EA-GMDH) combination of evolutionary algorithm, for the first time for cloud computing environment of host load prediction. By using the PSR method, the time series of one dimensional work load the form of time series in a multidimensional space, into the EA-GMDH neural network, the forecasting results through training and learning. Through the use of real Google cloud data center work load data by simulation experiment. This paper presents the EA-GMDH algorithm in average load forecasting and accurate load forecasting algorithm benchmark two aspects are better than the contrast Bayes, EMA, LWMA, LMA, AR, ANN and PP.4. proposed a new implementation of the cloud environment host load prediction method, using Autoencoder network and self encoding The Softmax classifier will return problem into a classification problem, again to enhance the load forecasting accuracy. Through an unsupervised learning method successfully the history window in the load data feature extraction. Using these extracted feature classifier, based on the successful prediction results of load data. Also, through the use of Google cloud the data center work load data as the data source for simulation. With the reference algorithm ANN, compared with Bayes and EA-GMDH algorithm, the proposed algorithm can obtain higher accuracy performance, to better realize the cloud environment for dynamic virtual machine precise balance to provide service and support, to reduce the overall system in the energy consumption to meet the requirements of QoS.

【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP308

【相似文献】