云计算环境能效建模与优化的关键技术研究

发布时间：2018-03-05 21:06

本文选题：云计算　切入点：能效　出处：《云南大学》2016年博士论文　论文类型：学位论文

【摘要】：高性能、低功耗且具有QOS保障的高能效云是云计算领域的一个热点研究问题。目前的研究大多是通过限定一个约束条件寻求另外指标的最优化来实现三者之间的折衷或均衡,现有的方法存在四个瓶颈性的问题：1)涉及能效的参数众多,综合能效指标的统一模型难以建立；2)在能效评价过程中,缺乏对云计算环境的能效优劣程度直观、定性的评估；3)资源分配粒度单一、固定,难以实时感知作业任务执行过程中能效的动态变化；4)能效参数间密切耦合,相互影响,相互制约,建模的误差和干扰因素的影响,导致能效值呈现不确定性,系统不具有鲁棒性。论文针对云环境下面向独立应用任务的能效指标之间的定量分析和均衡性问题进行了研究,同时针对大规模数据密集型应用任务的能效预测和解耦分析展开了研究,主要研究内容包括以下几个方面：(1)针对独立应用任务的QoS与系统能耗之间关系的定量分析问题,论文提出了一种单位能耗计算量的能源效率模型。首先将能耗作为一个关键因素引入到QoS指标中,基于能耗、性能和QoS之间的相互关系建立了一个系统能效模型,根据该模型所计算的能效值,采用演化博弈策略设计实现了满足不同服务需求的资源配置。针对不同数量的独立应用作业任务,所提出的能源效率模型可以较好刻地画性能、能耗和QoS三者间的相互关系。(2)针对不同应用任务类型性能、功耗和QoS保障三者之间的均衡性问题,论文提出一种基于QOS参数归约的加权能效模型。首先,把系统性能指标引入QOS,并将多个离散的QoS参数度量值归约到同一个量纲区域内,获得评价权重矩阵,求得用户最终的QoS评价值,以单位能耗所提供的整体QoS水平值作为能效值,并根据能效值将云数据中心进行了分级标识。通过在单机环境和云环境下实验测试,该模型能够较好的描述CPU密集型计算、I/O密集型计算和交互式任务的能效值,定性的评估云环境下的系统能效。(3)针对大规模数据密集、计算密集和分析密集等密集型计算任务,资源分配对系统能效的感知往往滞后的问题,论文建立了一种基于Lasso回归的能效预测模型,根据该模型提出了一种能效感知的、基于粒度空间寻优的资源分配方法。首先,在围绕待处理作业任务特点和虚拟资源特征的基础上,选用粒度合适的“粒”作为任务处理对象,针对不同粒度的“任务粒”在多维资源粒度空间以能效驱动的方式分配合理的“资源粒”,根据用户实时的需求变化进行有效的粒度层次快速切换。所提出的方法针对密集型任务在资源分配的过程中可以实时有效的保障用户的QoS,提高系统吞吐量,同时降低系统的能源消耗。(4)针对大规模密集型计算任务能效模型建模的误差和干扰因素的影响,导致系统能效值呈现出不确定性,系统不具有鲁棒性的问题,论文提出一种基于模糊解耦的能效优化方法。通过建立能效的模糊规则和隶属度函数,设计了一种多变量模糊解耦控制器,将MIMO耦合的云系统转化为几个互不干扰的SISO输出模型。所提出的方法可以明确影响系统能效的关键参数并及时定位系统能效瓶颈,同时在指定的不确定界的扰动下,仍能维持预期的能效值。
[Abstract]:High performance, a hot research issue in high energy efficiency and low power consumption with QOS cloud security computing. Most of the current study is to seek another index between optimization by defining a constraint conditions to achieve the three compromise or equilibrium, the present method has four bottleneck problems: 1) parameters involved the energy efficiency of many, it is difficult to establish a unified model of comprehensive energy efficiency index; 2) in energy efficiency in the process of evaluation, the lack of efficiency of the extent of cloud computing environment of the intuitive, qualitative evaluation; 3) resource allocation granularity is single, fixed task execution, sensing the dynamic change process of energy efficiency to real time; 4) energy efficiency parameters close coupling, mutual influence, mutual restraint, error and interference factor modeling, resulting in energy efficiency value is uncertain, the system does not have the robustness. The thesis focuses on the cloud environment to the independent application as below The quantitative analysis and the equilibrium problem between the energy efficiency index is studied, and the research of energy efficiency prediction and decoupling analysis for large-scale data intensive application tasks, the main contents include the following aspects: (1) the problem of quantitative analysis of the relationship between QoS and energy consumption of the system according to the independent application task, is proposed in this paper a model of energy efficiency calculation unit energy consumption. The energy consumption as a key factor is introduced to the QoS index, based on energy consumption, the relationship between performance and QoS established a system efficiency model, according to the calculation value of energy efficiency, using the evolutionary game strategy is designed to meet the needs of different service the allocation of resources. In view of the independent application tasks in different quantities, energy efficiency of the proposed model can better moment painting performance, mutual energy and QoS between the three . (2) according to the different types of application of task performance, the balance of power between the three and QoS security, this paper proposes a weighted QOS model parameters based on reduction efficiency. Firstly, the system performance index QOS is introduced, and a plurality of discrete QoS parameters to measure value reduction to the same dimension within the region and get the weights of the evaluation matrix, the QoS evaluation of the final value of the user, to provide the overall energy consumption per unit of QoS energy level value as the value, and according to the efficiency value of cloud data center to the classification identification. Through experiments in the stand-alone environment and cloud environment test, the model can describe CPU intensive computing, I/O intensive computing and interactive tasks. The value of energy efficiency, energy efficiency evaluation system under the cloud environment qualitative. (3) for large-scale data intensive computing intensive and intensive analysis and computation intensive tasks, resources distribution system efficiency Perceptions tend to lag problem, this thesis proposes a prediction model based on Lasso regression efficiency, according to the model proposed an energy aware resource allocation method, spatial optimization based on particle size. First of all, in around the processing based task characteristics and virtual resources characteristic, the suitable size of granule as the task object, according to the different size of "task grain" reasonable distribution of energy efficiency drive in the multi-dimensional resource granularity space mode of "resource grain", effectively fast switching granularity according to user needs change in real time. The proposed method for intensive tasks in the process of resource allocation in real time can be effectively protected users of QoS, improve the system throughput, and reduce system energy consumption. (4) for large scale intensive computing task efficiency modeling error and interference factors In effect, the energy efficiency of the system value is uncertain, the system does not have the robustness problem, this paper proposes a method of energy efficiency optimization based on fuzzy decoupling. Through the establishment of fuzzy rules and membership functions of the energy efficiency, design a multivariable fuzzy decoupling controller, cloud system MIMO coupling into the SISO output model do not interfere with each other. The proposed method can clear the key parameters influencing the system efficiency and timely positioning system efficiency bottleneck, at the same time in the specified uncertainties disturbances can still maintain the expected value of energy efficiency.

【学位授予单位】：云南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP3

【相似文献】