【作者】 张旭；

【导师】 尹浩；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2017， 博士

【摘要】 云计算作为新兴的计算模式,支持云资源的按需付费及云服务的弹性伸缩,使得全球云市场迅速扩张。随着物联网及超高清视频等新技术的出现,越来越多的云服务提供商开始在网络边缘部署云资源来提高用户访问云服务时的用户体验。本文主要研究如何以大数据的思想来优化云资源部署,提高用户访问云服务时的用户体验,降低云服务的部署成本,实现云服务的“提速降价”。为了实现云资源的高效部署,云服务提供商一方面需要在大规模场景下实现对网络的高效感知,为云资源部署提供数据支撑;另一方面需要在网络感知的前提下实现云基础设施的有效部署,包括计算、存储和带宽资源,为云服务提供基础设施支撑;最后在给定的基础设施上实现云服务高效部署,以云视频服务为例,优化用户观看视频用户体验的同时并降低视频的服务成本。然而,现有的网络感知方案难以同时兼顾测量准确性、测量成本以及网络友好性;现有基础设施部署方案受限于事先给定的部署候选点集,难以同时满足多种实际的部署需求;而现有的云视频部署方案难以实现视频云存储、云转码以及云分发的联合优化。为此,本文提出了数据驱动的云资源部署方案,主要包括基于广泛分布客户端的网络高效感知、数据驱动的云基础设施部署、和以用户为中心的云视频服务部署,主要创新点包括:首先,提出了基于广泛分布客户端的统计时延测量方法,在不影响用户使用客户端用户体验的前提下,可以高效地测量网络中任意两个IP之间以及任意两组IP之间的统计时延,无需在网络中部署大量额外测量节点,测量准确度高,网络友好性强。其次,提出了一套以用户为中心、以用户端性能数据等信息驱动的网络基础设施部署方法,突破了传统基础设施部署方法受限于事先给定候选点集的难题,得到了基础设施部署成本与用户体验等多因素之间的连续关系曲线,为云基础设施的按需部署提供真正有效的决策支持。最后,提出了视频云存储、云转码以及云分发的联合优化,基于对该联合优化问题的建模分析,设计了帕累托最优的单用户服务部署方案及基于贪婪选择的多用户服务部署方案,在保证用户满意率的前提下大幅度降低了云视频服务的部署成本。更多还原

【Abstract】 As a new computing paradigm,cloud computing suppots pay-as-you-go resource consumption and flexible expansion of cloud services,making the global cloud market expanding rapidly.With the emergence of new technologies such as Internet of Things and ultra-high-definition video,more and more cloud service providers are deploying cloud resources at the edge of the network to improve users’ Quality of Experience(QoE)when accessing cloud services.This paper focuses on how to optimize the cloud resource deployment with the idea of big data to realize the "boost speeds,cut prices" of online services.In order to realize the efficient deployment of cloud resources,the cloud service provider needs to measure the large-scale network with high accuracy to provide data support for the cloud resource deployment;deploy the cloud infrastructure effectively to provide infrastructure support for cloud services;and finally deploy cloud services(this paper takes video services in the cloud as the case study)on the given infrastruc-ture to optimize the user’s QoE while reducing the cost of video services.However,the existing network awareness scheme is difficult to take into account both measure-ment accuracy,measurement cost and network friendliness.The existing infrastructure deployment scheme is limited by the pre-given deployment candidate set,and it is dif-ficult to meet a variety of actual deployments requirements at the same time;and the existing cloud video optimization strategy is difficult to achieve the joint optimization of video cloud storage,cloud transcoding and cloud distribution.In this paper,we design a data-driven video cloud resource deployment scheme,which mainly includes the efficient network awareness,data-driven cloud infrastructure deployment,and user-centric cloud video service deployment based on widely distributed client.The research work includes:First of all,we proposed a network awareness scheme based on the widely distributed clients in the network.The scheme can achieve high accurate real-time delay between arbitrary two hosts and the statistical delay between two IP host sets without sacrific-ing users’ QoE.Moreover,the scheme do not need to deploy a large number of extra measurement nodes in the network,and is network-friendly.Secondly,we proposed a user-centric framework for network infrastructure deploy-ment,which can meet the various requirements of cloud service providers with flexible infrastructure deployment.Moreover,the infrastructure deployment performance and cost are no longer limited to the pre-given candidate set.Based on the continuous relationship between the infrastructure deployment costs and the user experience,the framework can provide an effective decision support for on-demand deployment of cloud infrastructure.Finally,we proposed the joint optimization of video cloud storage,cloud transcoding and cloud distribution.Based on the modeling and analysis of the joint optimization prob-lem,we designed a Pareto-optimal deployment scheme for single-user and a greedy-based service deployment scheme for multi-user.The schemes can reduce the video service deployment cost significantly when ensuring the user satisfaction rate simultaneously.更多还原