云视频服务系统的资源配置和调度优化

发布时间：2018-09-14 08:23

【摘要】：随着高速互联网的广泛普及和多媒体技术的飞速发展,通过互联网在线观看视频已成为最受欢迎的互联网应用之一。传统上,视频服务提供商主要基于内容分发网络和对等网络的架构来构建大规模视频服务系统。然而,内容分发网络半静态的资源配置方式使得系统可扩展性较差,资源利用率较低和服务成本高昂,难以应对不断增长的用户请求。而对等网络依赖用户上传的互相配合,难以保证服务的可靠性,无法保障用户对视频观看体验的更高要求。相对于这些传统技术,新兴的云计算提供了一种可靠、弹性和低成本的资源配置方式,这为视频服务商带来了新的解决方案。本文考虑利用云计算平台构建新一代的大规模视频服务系统。视频服务提供商通过按需动态调整配置的云计算资源,来应对高度异构和动态的视频需求并同时降低运营成本。本文针对基于云的大规模视频服务系统中面临的动态资源配置、请求调度、内容部署和网络管理等问题进行了深入研究,通过建立数学模型来描述系统,在模型下利用相关优化理论推导出系统的最优控制策略。具体的研究工作概括如下：1)考虑云计算服务商以虚拟机为单位向视频服务商提供计算资源,并提供多种租用定价方案。视频服务商通过按需调整所租用的虚拟机数量,来应对用户动态的视频请求并降低运营成本。本论文将上述问题建模成满足服务质量约束下系统应配置的虚拟机最小数量问题和多种定价模型下各种类型虚拟机的最优采购方案问题。将服务质量约束表示成系统过载的概率需要在设定的阈值范围内,并利用大偏差原理估计过载概率,进而基于在线测量得出虚拟机的最优配置数量。然后通过分析比较多种定价模型,利用整数规划求出每个时隙虚拟机的最优租用方案,并提出了一个动态调整长期预留型虚拟机数目的优化策略。通过实验仿真验证了算法的有效性。2)研究了视频服务商利用多个分布各地的云数据中心来构建大规模视频服务系统向多地区的用户进行服务的优化问题。为了提升用户观看体验质量以最大化系统收益并同时降低运营成本,需要以一定策略调度并服务来自多地区的用户请求,并同时不断按需调整内容部署和资源配置。本文利用Markov决策过程来建模描述多数据中心云视频服务系统的动态运行过程。利用平均性能准则来反映系统长期运行的平均利润,其中利润为系统收益减去运营成本,系统收益通过一个效用函数反映了用户的观看体验质量。通过寻求一个动态的请求调度、视频部署和资源配置联合优化策略来最大化平均利润。利用性能灵敏度分析,本文设计了一个基于样本轨道的策略迭代算法来找到最优策略,并证明了算法的最优性。接着从实际系统运行的角度对该算法进行了详细的分析,并通过一系列仿真实验验证了本文的算法有效性。3)考虑先进的软件定义网络(Software Defined Network, SDN)技术和云计算技术的结合。本文提出了一个SDN功能化的媒体云架构,云数据中心之间通过SDN功能化的网络互相连接。针对网络拥塞对视频服务传输时延以及对用户观看体验的影响,通过将路由策略与请求调度和资源分配策略联合优化,以降低移动云视频服务商的运营成本并提升视频播放质量。借助效用函数来反映视频播放质量的两个方面：用户的满意度和平均延迟。系统目标函数定义为最大化视频服务商的利润(即效用收益减去运行成本),并通过一个统一的流量管理模型合并用户请求调度和视频流路径选择决策,将联合优化问题描述成一个混合整数规划问题。基于对偶分解和次梯度的方法,提出了一个迭代算法来求解混合整数规划问题,该算法可以有效地得到内容部署和流量管理的最优决策。以实际环境的参数设置进行仿真实验,验证了该算法的有效性。实验结果表明,本文提出的算法可以同时实现优化运营成本和用户体验之间的良好平衡。
[Abstract]:With the popularity of high-speed Internet and the rapid development of multimedia technology, watching video online over the Internet has become one of the most popular Internet applications. Traditionally, video service providers mainly build large-scale video service systems based on the architecture of content distribution networks and peer-to-peer networks. Static resource allocation makes the system less scalable, lower resource utilization and high service cost, which makes it difficult to cope with increasing user requests. Peer-to-peer networks rely on the cooperation of user uploads, which makes it difficult to guarantee the reliability of the service and the higher requirements of users for video viewing experience. Technology, emerging cloud computing provides a reliable, flexible and low-cost way of resource allocation, which brings new solutions for video service providers. This paper considers building a new generation of large-scale video service systems using cloud computing platforms. In this paper, the dynamic resource allocation, request scheduling, content deployment and network management problems in cloud-based large-scale video service systems are studied in depth. A mathematical model is established to describe the system and the related optimization theory is used to derive the model. The detailed research work can be summarized as follows: 1) Considering that cloud computing service providers provide computing resources to video service providers in virtual machines, and provide a variety of rental pricing schemes. In this paper, the above-mentioned problems are modeled as the minimum number of virtual machines to be configured to satisfy the quality of service constraints and the optimal procurement scheme for various types of virtual machines under various pricing models. Based on the overload probability, the optimal number of virtual machines is obtained by on-line measurement. Then the optimal rental scheme of each slot virtual machine is obtained by analyzing and comparing various pricing models, and an optimization strategy is proposed to dynamically adjust the number of long-term reserved virtual machines. Effectiveness. 2) This paper studies the optimization problem of video service providers using cloud data centers to construct large-scale video service systems to serve multi-regional users. In order to improve the quality of user viewing experience to maximize system benefits and reduce operating costs, it is necessary to schedule and serve multi-regional users with a certain strategy. In this paper, Markov decision process is used to model and describe the dynamic operation process of cloud video service system in multi-data center. The average performance criterion is used to reflect the average profit of long-term operation of the system, in which the profit is the system income minus the operation cost and the system income. A utility function is used to reflect the quality of viewing experience of the user.A joint optimization strategy of dynamic request scheduling, video deployment and resource allocation is sought to maximize the average profit.Using performance sensitivity analysis, a strategy iteration algorithm based on sample trajectory is designed to find the optimal strategy, and the algorithm is proved. Then the algorithm is analyzed in detail from the point of view of the actual system operation, and the effectiveness of the algorithm is verified by a series of simulation experiments. 3) Considering the combination of advanced software Defined Network (SDN) technology and cloud computing technology, this paper proposes a SDN functionalized media cloud framework. Cloud data centers are interconnected via SDN functionalized networks. In view of the impact of network congestion on the transmission delay of video services and the viewing experience of users, the routing strategy is optimized in combination with the request scheduling and resource allocation strategy to reduce the operating costs of mobile cloud video service providers and improve the quality of video playback. Utility function reflects two aspects of video playback quality: user satisfaction and average delay. The system objective function is defined as maximizing the profit of video service providers (i.e. utility revenue minus running cost). A unified traffic management model is used to merge user request scheduling and video stream path selection decisions, and the joint optimization is achieved. The problem is described as a mixed integer programming problem. Based on the dual decomposition and subgradient method, an iterative algorithm is proposed to solve the mixed integer programming problem. The algorithm can effectively obtain the optimal decision-making of content deployment and traffic management. Experimental results show that the proposed algorithm can achieve a good balance between optimal operating costs and user experience.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP393.09

【相似文献】