云计算性能与节能的动态优化研究

发布时间：2019-02-14 22:32

【摘要】：随着云计算(CC, Cloud Computing)勺蓬勃发展,云数据中心高能耗、高碳排放的问题日益凸显,给云服务提供商带来高额运营成本的同时,严重制约了云计算的可持续发展。云计算应用领域的不断拓展使其服务对象已由传统的桌面用户群渗透到移动用户群,催生了移动云计算(MC2, Mobile Cloud Computing)这一新兴计算模式。MC2通过移动互联网连接移动设备端与云端,对端到端数据传输的能效提出了较高的要求。本文围绕CC和MC2的性能与节能优化展开研究,运用动态优化方法构建理论分析模型,设计在线控制算法,优化系统的能耗和性能。论文的研究内容和成果包括： (1)数据中心计算资源自配置的性能与节能优化。首先,运用马尔科夫决策过程(MDP, Markov Decision Process)理论构建资源自配置问题的动态优化模型；然后,鉴于外部环境模型的未知性,综合运用强化学习和近似动态规划方法,提出了一种计算资源自配置算法RASA, RASA算法利用服务器CPU的动态频率调节机制,动态匹配资源分配量与系统负载,优化系统能耗和性能；仿真实验验证了RASA算法的有效性。 (2)分布式SaaS云请求路由与虚拟机调度的节能优化。首先,构建分布式SaaS云成本与性能管理问题的动态优化模型,目标是在保证应用请求队列稳定性的前提下,最小化时间平均(Time Average)能源成本、碳税成本和带宽租用成本；然后,运用Lyapunov随机优化方法,提出了一种分布式的在线调度算法GREEN,在运营成本最优性与队列稳定性之间实现平衡控制；最后,设计基于真实数据集的仿真实验,验证GREEN算法在非稳态环境下的有效性。 (3)MC2链路选择与传输调度的性能与节能优化。首先,运用MDP理论构建端到端上下行数据传输问题的动态优化模型；然后,提出了一种基于近似动态规划的在线学习算法eLean,该算法利用不同链路的能效差异性和部分移动应用的延迟容忍特性,通过动态的链路选择与数据传输调度,优化移动设备能耗和吞吐量；最后,设计仿真实验对eLean算法的有效性进行了验证。
[Abstract]:With the rapid development of cloud computing (CC, Cloud Computing), the problem of high energy consumption and high carbon emissions in cloud data centers is becoming increasingly prominent, which brings high operating costs to cloud service providers and seriously restricts the sustainable development of cloud computing. With the continuous expansion of cloud computing application field, the traditional desktop user group has penetrated into the mobile user group, which has given birth to the mobile cloud computing (MC2,). Mobile Cloud Computing) is a new computing mode. MC2 demands high efficiency of end-to-end data transmission by connecting mobile device and cloud via mobile Internet. This paper focuses on the performance and energy saving optimization of CC and MC2. The dynamic optimization method is used to construct the theoretical analysis model and design the on-line control algorithm to optimize the energy consumption and performance of the system. The research contents and achievements are as follows: (1) performance and energy saving optimization of data center computing resource self-configuration. Firstly, the (MDP, Markov Decision Process) theory of Markov decision process is used to construct the dynamic optimization model of resource self-allocation problem. Then, in view of the uncertainty of the external environment model, a computational resource self-configuration algorithm (RASA, RASA) is proposed to utilize the dynamic frequency regulation mechanism of server CPU by using reinforcement learning and approximate dynamic programming. Dynamically match resource allocation with system load, optimize system energy consumption and performance; Simulation results show the effectiveness of RASA algorithm. (2) Energy saving optimization of distributed SaaS cloud request routing and virtual machine scheduling. Firstly, the dynamic optimization model of distributed SaaS cloud cost and performance management is constructed. The objective is to minimize the time average (Time Average) energy cost, carbon tax cost and bandwidth rental cost under the premise of ensuring the stability of application request queue. Then, using Lyapunov stochastic optimization method, a distributed online scheduling algorithm, GREEN, is proposed to achieve balance control between operational cost optimality and queue stability. Finally, a simulation experiment based on real data set is designed to verify the effectiveness of GREEN algorithm in unsteady environment. (3) performance and energy saving optimization of MC2 link selection and transmission scheduling. Firstly, the dynamic optimization model of end-to-end uplink and downlink data transmission is constructed by using MDP theory. Then, an online learning algorithm based on approximate dynamic programming (eLean,) is proposed, which makes use of the difference of energy efficiency of different links and the delay tolerance of some mobile applications, through dynamic link selection and data transmission scheduling. Optimize energy consumption and throughput of mobile devices; Finally, simulation experiments are designed to verify the effectiveness of the eLean algorithm.
【学位授予单位】：北京科技大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP308

【参考文献】