云计算环境下的仿生自主监控系统和多指标均衡调度机制的研究

发布时间：2018-03-10 10:25

本文选题：仿生自主神经系统　切入点：监控　出处：《电子科技大学》2017年博士论文　论文类型：学位论文

【摘要】：随着云计算技术的快速发展和应用普及,使得以往分散的资源再次呈现出集中的趋势,这使得云计算环境的规模也相应的变的越来越大,复杂性也相应提高。如何保障云计算环境的可靠运行,如何有效的评价云计算环境的服务性能,又该如何应对云计算环境庞大的能耗问题,以及研究有效处理用户服务请求和资源优化供给之间的均衡优化调度,都成为迫切的需求。针对于此,必须对可靠计算、高性能计算和节能计算技术进行深入研究,综合可靠性、性能和能耗等多指标,建立多维关联的评估模型,并基于综合评估模型研究面向多目标的均衡调度机制。为了保障可靠性,监控是重要的保障手段,但现有监控缺乏自主自治能力,系统架构也不太适合云计算环境;为了进行性能评估,必须充分考虑云服务的特征才能建立合适的性能模型,并有效考虑可靠性对性能的影响才能更准确的评估云服务的性能;为了节能高效计算,传统方法仅将能耗和性能作为相互制约的因素,而忽略了可靠性。在已有的研究中,将可靠性、性能和能耗分离割裂的评估建模,也导致了基于相关模型的优化调度带有局限性。针对以上问题,本文主要研究了面向云计算环境的仿生自主监控系统和多指标关联的均衡调度机制,主要研究内容及贡献体现在四个方面,如下:1)针对传统监控系统扩展性差、缺乏自主性,以及未能充分考虑云计算虚拟化和动态弹性等问题,设计了一种具备仿生自主神经系统特征的云监控系统。该云监控系统具备“人机交互层——中枢神经系统层——周围神经系统层——神经元层——神经突轴层”五层的分层架构,融入自组织、自诊断、自修复和自行为等自主特性。新的架构设计,采用周围神经和神经元的结构能充分反映虚拟资源的宿主特性,同时设计周围神经系统内部的自治特性,即能完成自治任务,也可有效减轻中枢压力;逐级扩展的树型结构赋予系统良好的扩展性。将自主特性融入云监控系统,有效提升了监控系统对规模复杂云计算环境的适应性和健壮性,并利于自动扩展。通过自组织功能有效解决了规模虚拟化资源因动态变化(加入、退出等)带来的监控难和部署繁琐问题;通过自诊断和自修复提升云监控系统自身的稳定可靠性;通过自行为模块的设计,不仅增强了系统的扩展功能,也减少了不必要的数据传输,降低了对系统的开销。2)针对传统性能评估忽略可靠性的问题,提出了一个面向云服务特征的关联“可靠性——性能”的性靠度模型。云计算环境具有异构性、虚拟化和按需服务特征,其关联性能(性靠度)评估模型也因此不同。考虑虚拟资源的失效和修复行为,运用Markov链对其可靠性进行评估,提出采用通用生成函数(Universal Generating Function,UGF)分析异构云资源池可靠性的方法;考虑按需服务特性,分析1:1和1:D映射模式下云服务的不同性能,建立反映请求生灭过程的云服务系统性能模型;考虑资源可靠随机性对服务性能随机性的影响,通过Baysian方法将可靠性、性能模型关联,建立新的面向云计算特征的性靠度模型。实验结果表明,该二维关联建模方法,能更准确的评价真实云计算环境的服务性能。其意义还在于,提供了一种通用的分析不同的性靠度关联的方法。3)针对传统评估方法将可靠性、性能和能耗指标割裂开来的不足,以及云计算环境所面临的严重能耗的现实问题,提出了一种关联“可靠性——性能——能耗”的多维关联的层级建模方法,建立综合性评估指标。该层级建模的思想是:提出和求解一个复杂系统的全局模型往往较为复杂,本文采取将全局模型分解为覆盖交互因子的多个子模型的方法,然后通过对子模型的迭代求解可以获得全局模型,以顺利的完成系统的整体评估。在云计算环境中,多维层次模型的关联交互存在一个共同的条件参数——“可用资源随机数量”,基于该条件参数所建立的各个子模型,最后可通过Bayesian途径,移除条件参数,最终整合各个子模型,并获得多指标关联的综合性评价指标——期望性能和期望能耗。采用该方法,本文将基于真实采样数据进行统计分析所获得的能耗模型与可靠性和性能关联,通过实验验证了多维关联的能耗评估的有效性。4)针对云计算环境所面临的复杂供需制约调度需求,及现有层次化调度缺乏自主性和优化调度目标单一等问题,提出了一种分层优化调度模型和多指标联合优化的调度策略。其具体工作机制为:全局代理负责全局的任务分发,体现集中管控;局部代理实现区域自治,通过本地自主计算形成反映任务不同分配情况的局部资源优化的“供需地图”,在多指标联合的利润目标优化下实现资源的合理调度供给以匹配全局代理动态变化的任务需求。在局部优化的基础上,利用反映多维关联的优化函数,设计具体的遗传算法进行问题的求解,实现多目标联合优化下的全局最优。实验结果表明该分层优化调度机制能够提高搜索最优解的效率,可以获得综合的优化效果。
[Abstract]:With the rapid development and popularization of the application of cloud computing technology, which makes distributed resources once again showing the concentration trend, which makes the scale of cloud computing environment also become more and more big, the complexity is increased accordingly. How to ensure the reliable operation of the cloud computing environment, such as service performance evaluation of any effective cloud computing environment the problem of energy consumption, environmental computation and how to deal with the cloud, and study the effective treatment of equilibrium between user service request scheduling and optimization of resource supply, has become the urgent demand. According to this, must be reliable computing, high performance computing and saving computing technology in-depth research, comprehensive reliability, performance and energy consumption etc. multi index evaluation model, multidimensional association, and the scheduling mechanism for multi-objective comprehensive evaluation model based on. In order to ensure the reliability, monitoring is an important guarantee But the existing monitoring means, lack of autonomy ability, system architecture is not suitable for cloud computing environment; in order to evaluate the performance, must fully consider the characteristics of cloud services to establish suitable performance models, and consider the performance reliability of the performance impact of cloud services to assess more accurately; in order to energy efficient computing, traditional method only the energy consumption and performance as factors restrict each other, while ignoring the reliability. In the previous study, the reliability evaluation model of separation between performance and energy consumption, also led to the optimization of scheduling based on correlation model with limitations. In view of the above problems, this paper mainly studies the Bionics for cloud computing scheduling mechanism independent monitoring system and related indicators, the main research contents and contributions are in four aspects: 1) the traditional monitoring system scalability, lack of self The Lord, and do not consider the virtual and dynamic elastic problems such as cloud computing, design a cloud monitoring system with bionic autonomic nervous system characteristics. The cloud monitoring system with hierarchical architecture of human-computer interaction layer of central nervous system, peripheral nervous system neurons - layer - layer neural axonal layer "the five layer, into self organization, self diagnosis, self repair and self independent characteristics. The new architecture design, the structure and characteristics of host peripheral nerve neurons can fully reflect the virtual resource, and design of autonomous internal characteristics of the peripheral nervous system, which can accomplish autonomous tasks, can effectively alleviate the central pressure; the tree structure gradually extended to the good scalability of the system. The independent characteristics into the cloud monitoring system, improve the monitoring system of computing environment adaptability and robustness of complex cloud scale And, to automatically expand. Through the self-organization function effectively solves the scale of virtualized resources due to dynamic changes (join, exit) caused by monitoring difficult and complicated problems through deployment; self diagnosis and self repair to enhance reliability of cloud monitoring system's stability; by module design, not only enhances the expansion function of the system also, reduce the unnecessary data transmission, reduces the system overhead of.2) to assess the reliability of the traditional ignore performance, proposed a feature oriented Cloud Service Association "reliability performance" of reliability model. Cloud computing environment is heterogeneous, virtualization and on-demand service features. The association performance evaluation model (of reliability) are so different. Considering the virtual resource failure and repair, the use of Markov chain to assess its reliability, the general production function (Universal Ge Nerating Function, UGF) analysis method of heterogeneous cloud resource pool reliability; according to need service characteristics, analysis of different performance of cloud services 1:1 and 1:D mapping mode, the establishment of performance reflect the request of cloud service system model of the birth and death process; considering the influence on the performance of reliable resources random random, through the Baysian method to reliability the performance model, association, build cloud computing features of the new reliability model. The experimental results show that the two-dimensional correlation modeling method can more accurately evaluate the real cloud service performance computing environment. Its significance lies, provides a general analysis of different methods on the degree of the associated.3) for the traditional evaluation method of reliability, low performance and energy consumption index separated, serious problems of energy consumption and cloud computing environment, proposes a reliable performance of association "- - hierarchical modeling method of multidimensional association energy consumption ", the establishment of a comprehensive evaluation index. The hierarchical modeling idea is put forward: the global model and solving a complex system is often complex, this paper adopts the global model is decomposed into several sub models covering the interaction factor, and then through the iterative sub model can be obtained the global model, in order to complete the whole evaluation system smoothly. In the cloud computing environment, related multidimensional hierarchical model interaction there is a common condition -" random number of available resources, each sub model parameters based on the condition, finally through the Bayesian pathway, remove the condition parameters, final integration sub model and multi index comprehensive evaluation index association -- expected performance and expected energy consumption. Using this method, this paper will be based on the real data sampling system The analysis model of energy consumption and reliability and performance, through experimental verification of the.4 effectiveness evaluation of the Energy Association) for cloud computing complex supply and demand facing the environment and the existing scheduling needs, hierarchical scheduling scheduling problems such as lack of autonomy and single objective optimization, this paper proposed a novel scheduling model and multi joint optimization index hierarchical optimization scheduling strategy. The specific mechanism for global agency is responsible for the global task distribution, reflect centralized management; implementation of regional autonomy of local agents, through local autonomic computing form reflect the distribution of the different tasks of local resource optimization "supply and demand map", the supply of resources in a reasonable scheduling optimization combined with profit the parameters to match the dynamic changes of the global agency requirements. Based on local optimization, the optimization of the letter reflects the Multidimensional Association Number, we design specific genetic algorithm to solve the problem, and achieve the global optimization under multi-objective joint optimization. The experimental results show that the hierarchical optimization scheduling mechanism can improve the efficiency of search optimal solution, and achieve comprehensive optimization results.

【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP393.09

【参考文献】