【作者】 敬思远；

【导师】 佘堃；

【作者基本信息】 电子科技大学 ， 计算机应用技术， 2013， 博士

【摘要】 随着Internet的繁荣以及云计算技术的发展，数据中心正变得空前重要。但是当前数据中心动辄装备十万、百万规模的服务器以及相应的其它IT基础设施，使资源管理变得非常困难。此外，海量的基础设施时刻消耗巨大的电能，这不仅增加了运营商的运营成本，而且导致大量的二氧化碳排放。数据中心资源虚拟化、资源管理动态化、弹性化、自动化、节能化成为业界一致认可的发展趋势。面向节能的自适应资源管理是当前研究绿色虚拟数据中心的热点之一。论文相关工作以此为依据，首先分析和总结了当前数据中心中基于资源管理的主要节能技术，然后围绕虚拟数据中心资源管理中的若干关键问题进行了深入研究。论文的主要工作包括：1.提出了一种基于管理员资源选择偏好的虚拟数据中心资源规划方法。首先基于粒度粗糙理论，提出了支持偏好语义表述的偏好粒子模型以及纵向聚合运算操作符和横向融合运算操作符，建立了一种基于粗糙粒度表示的用户偏好模型，解决了当前用户偏好模型不支持语义表述以及不能处理不确定性偏好的问题。其次，将该模型运用于描述数据中心资源规划过程中的管理员的资源选择偏好。同时，基于不同应用的负载峰值会出现在一天中的不同时段的原理，提出了基于应用负载分时分析的建模方法，在保证不违背服务等级协议（Service LevelAgreement, SLA）的前提下能够使资源利用率进一步提升。考虑了虚拟机（VirtualMachine, VM）之间的联系性和互斥性，VM与服务器之间的兼容性，提出了5项资源规划原则。将资源规划问题看作是有约束的多维装箱问题，提出了基于分组遗传算法（Grouping GeneticAlgorithm，GGA）的智能优化算法搜索全局最优解。最后通过仿真实验对模型进行了验证。2.针对当前虚拟数据中心中面向节能的动态资源优化方法仅考虑资源利用率最大化，而忽略了VM位置变化（包括VM启动、迁移和回收）带来的系统开销问题，提出一种能同时减少系统能耗以及VM位置变化的动态资源优化新方法。该方法采用基于周期性控制回路的资源管理框架，将动态优化分配问题表述为一个双目标组合优化问题，并提出了一种基于网络流理论和迭代优化的近似优化算法NFT-DRP(Network Flow Theory based Dynamical Resource Provision)。实验结果表明，新方法在降低系统能耗方面与现有方法相比略有提升，而在减少VM位置变化方面则有较大改进。3.针对当前虚拟数据中心面向节能的VM动态整合研究仅考虑服务器能耗，而忽略了网络设备能耗的现状，提出了一种同时降低服务器和网络设备能耗的新方法。该方法通过感知数据中心的网络拓扑结构，采用一种混合粒子群优化算法HPSO-NA(Hybrid Particle Swarm Optimization-Network Aware)来计算最优的VM整合方案，使运行的服务器和网络设备最少以实现最小化能耗。实验结果表明该方法能进一步降低系统整体能耗。4.针对当前主要的热点检测模型均只能基于传统资源（CPU，内存，网络带宽）进行决策的现状，提出了一种新的基于多条件决策的虚拟数据中心热点检测模型。该模型通过将不同量纲的系统特征值纳入一个统一决策模型，实现对系统中的服务器和VM进行评分。采用了阈值法，通过设置热点阈值(Hotspot DetectingThreshold，HDT)对系统热点进行检测，当服务器评分超过HDT时，从该服务器上选择VM进行迁移以消除热点；当VM的评分超过HDT时，则首先判断是否可以通过本地资源再分配消除热点，否则对该VM进行迁移。实验结果证明了该模型的有效性。5.针对当前CMOS多核嵌入式处理器片上仅提供全局动态电压缩放（Dynamic Voltage Scaling，DVS）支持以及亚纳米时代后CMOS处理器泄漏功耗不可忽视的现状，提出了一种新的多核嵌入式环境中的硬实时任务感功调度算法GRR&CS。算法通过基于贪心法的静态任务划分，基于全局资源回收利用和任务迁移的动态负载均衡，以及动态核缩放三个步骤实现整体能耗的降低，并同时保证实时任务的可调度性约束。实验表明，提出的算法相比较现有算法多节省14.8%-41.2%的能耗。更多还原

【Abstract】 With the prosperity of Internet and development of cloud computing technology,data center becomes more and more important. Currently data center hosts overhundreds of thousands of servers as well as other IT infrastructure devices which notonly makes managerial work more difficult, but also consumes huge amount of energyat the same time. Furthermore, this not only reduces the profit margin of serviceprovider, but also leads to high carbon emission. A unanimously approved trend is thatdata center resource virtualization and resource management has become dynamic,flexible, automatic and energy efficient. One of the hot issues in this research domain isenergy-efficient adaptive resource management. Based on this proposal, this thesisanalyzes and concludes the state-of-art techniques and research on energy efficientresource management in data center at first, then do depth study of some key issuesaround resource management in virtualized data center. The main contributions in thisthesis include:1. An administrator’s preference based method for capacity planning in virtualizedenterprise data center is proposed. First, based on granular rough theory, a family ofatomic preference granules supporting semantic alongwith two operators, namedvertical aggregation operator and horizon combination operator are proposed, and anew preference model, which makes use of rough granules to describe user preference,is also built. The new model can address the said issue that existing scheme cannot dealwith using semantic preference representation and uncertain data. Second, thepreference model is applied to describe the administrator’s preference of resourceselection. Meanwhile, due to the peak-valley of the workload of applications is indifferent time, a way to problem modeling is proposed based on workload timesharinganalysis, which can achieve maximal resource utilization on the premise of keeping therequired QoS. The associated services and the mutually exclusive services, thecompatibility between services and servers are taken into account and five principles ofconsolidation are proposed. The establishing model is regarded as a multi-dimension binpacking problem with several constraints and a GGA-based algorithm is proposed to search for the global optimal solution. Finally, the proposed method is proved with thehelp of experiments.2. Since existing methods of energy efficient dynamic resource optimization indata center just maximize resource utilization without considering the overhead of VM(Virtual Machine) placement change (e.g. VM deployment, start, migration and reclaim),we propose a new dynamical resource optimization method which can minimize theenergy consumption and VM placement change at the same time. The new methodadopts control loop based resource management frame work, and regards dynamicalresource optimization as a dual-objective combinational optimization problem, andmoreover it designs a network-flow-theory and iterative optimization basedapproximate algorithm, called NFT-DRP, to solve it. The experimental results whencompared to existing work show that, the proposed method can slightly decrease theenergy consumption but greatly decrease the number of VM placement change.3. Since current researches only consider part of energy consumed by servers, anovel method is proposed to reduce energy consumption from both servers and networkdevices. The proposal achieves minimum energy consumption by minimizing the activeservers and network devices through analyzing the net topology. Moreover, the workformally models the problem and designs a hybrid particle swarm algorithm calledHPSO-NA to implement virtual machine consolidation. The experimental results showthat the proposed method can effectively reduce the energy consumption.4. Since current hotspot detecting algorithms work based on traditional resource(i.e. CPU, memory, bandwidth, etc.), a new multi-criteria decision making based modelis proposed to detect hotspot and make decisions of VM scheduling in virtualized datacenter. By incorporating system eigenvalue with different dimensions into the unifieddecision model, it can score the servers and VM in system. Moreover, a HDT ispredefined and applied to detect both server hotspot and VM hotspot. If the server’sscore is higher than HDT, we select suitable VM and shift them to remove hotspot.When VM’s score is higher than HDT, the VM’s resource is reallocated to removehotspot if it is possible; else the VM would be migrated to other server. Experimentalresults prove the efficiency of the model.5. Since that CMOS multi-core embedded processor only provides global DVS andits leakage power is negligible, this work proposes a new power-aware scheduling algorithm for hard real-time tasks in multi-core embedded environment, calledGRR&CS. The power saving is achieved by three steps including greedy-based statictasks partition, and global resource reclamation based dynamic load balance anddynamic core scaling. The algorithm also keeps the schedulability of tasks. Experimentsshow that the proposed algorithm saves14.8%-41.2%energy more than other existingworks.更多还原