【作者】 王博；

【导师】 任丰原；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2019， 博士

【摘要】 近年来,点播和直播等视频业务发展迅猛。为了满足用户对体验质量(QoE)越来越高的要求,目前的服务提供商主要使用HTTP自适应视频流(标准化为DASH)来传输视频。在DASH中,视频被切成等时长的块并编码成多种码率。客户端的码率自适应(ABR)算法根据系统状态(吞吐量和缓存量)以优化QoE为目标来动态调节码率。点播和直播对于QoE的要求不同。点播要求高码率、少卡顿、少码率切换,而直播要求低时延、少卡顿、高码率。选出一个恰当的码率是困难的:(1)吞吐量和缓存量在不稳定的网络中变化剧烈,吞吐量难以被准确预测,且码率易被切换;(2)相同码率的视频块有不同的块大小;(3)为实现QoE优化目标,直播ABR算法需要将缓存量控制在一个低但不为零的水平,这在不稳定的网络中难以实现。本文分别对点播流和直播流的码率自适应问题进行了如下四方面的研究。(1)提出了基于多步预测减少码率切换的点播算法。已有算法根据瞬时系统状态来选择码率,会引起频繁的码率切换。为了提升码率平滑性,本文基于多步预测设计了MSPC算法。实验结果表明,MSPC能够使码率仅适应网络的持续变化而忽略其短暂抖动,可以有效地减少码率切换。(2)分析并解决了点播中块大小波动的影响。本文通过实验观测分析了块大小的波动特性及其对QoE的影响。为了解决其引起的卡顿问题,本文建议使用实际块大小来选择码率,并设计了解决方案。实验结果证明了算法的有效性。(3)提出了基于优化控制增强鲁棒性的点播算法。目前的ABR算法只以优化QoE为目标而没有对缓存量进行控制,因此缓存量抖动大,吞吐量预测误差易引起卡顿。为了抑制预测误差的影响,本文将QoE优化与缓存控制相结合,并设计了基于优化控制的算法RBC。仿真结果表明RBC能够有效地减少卡顿时间。(4)提出了基于双馈联合方法精准控制缓存量的直播算法。为了提高控制能力,本文提出了将前馈控制与反馈控制相结合的双馈控制方法。实验结果表明,所设计的HCA算法能够将缓存量准确地控制在目标水平,能够同时实现低时延、少卡顿、高码率。更多还原

【Abstract】 Video traffic is prominent in recent years,including video-on-demand(Vo D)streaming and live streaming.To meet the increasing user demands on quality-of-experience(QoE),most content providers use HTTP adaptive streaming(standardized as DASH)to deliver video.In DASH,video is split into chunks encoded into multiple bitrate levels.Client-side adaptive bitrate(ABR)algorithm dynamically adjusts the bitrate at chunk boundaries.The bitrate is selected based on throughput prediction and buffer occupancy with the goal of maximizing QoE.For Vo D streaming,high QoE means high bitrate,low rebuffering and low bitrate switch.For live streaming,high QoE means low delay,low rebuffering and high bitrate.However,picking a correct bitrate is challenging since(1)in the highly variable network conditions where both the throughput and buffer occupancy change rapidly,the video bitrate can be switched frequently,and it is hard to obtain an accurate throughput prediction;(2)the video chunks at the same quality have various chunk sizes;(3)Optimizing QoE for live streaming requires the ABR algorithm to control the buffer occupancy at a small but non-empty level,which is challenging in the highly variable network.This thesis investigates the bitrate adaptation issue for both two kinds of video streaming in detail.(1)We propose a multi-step prediction based scheme to reduce bitrate switches.Existing ABR algorithms rely on instantaneous system state to select bitrate,which can induce serious bitrate switches in unstable networks.To improve the bitrate smoothness,we propose to use multi-step prediction to make bitrate selection and develop the MSPC scheme.Experiment results show that MSPC is able to effectively reduce the bitrate switches,with the video bitrates only adapting to sustainable network changes while ignoring the momentary fluctuations.(2)We analyze the characteristics of chunk size variation and its impact on QoE,and improve the bitrate adaptation to eliminate the negative impact.Through measurement experiments,we analyze the characteristics of chunk size variation and observe that it increases the risk of rebuffering.To eliminate this risk,we propose that the bitrate decision should be made based on the actual chunk size rather than the nominal bitrate.Experiment results demonstrate the effectiveness of the improved ABR algorithm.(3)We propose RBC algorithm by combining QoE optimization and buffer control to improve the robustness against throughput prediction errors.We observe that the sensitivity of existing ABR algorithms to throughput prediction error is resulted by the lack of buffer control.Since the bitrate selection only aims to optimize QoE without concerning the buffer dynamic,the buffer occupancy varies widely at great risk of rebuffering.To restrains the rebuffering problem induced by throughput prediction errors,we propose the RBC algorithm which controls the buffer occupancy within a safe range while making QoE optimization.By evaluation experiments,it shows that RBC can largely reduce the rebuffering time.(4)We use the hybrid of feedback control and feedforward control to develop HCA algorithm for live streaming which has strong control capability.Optimizing the QoE for low delay live streaming can be realized by controlling the buffer occupancy at a small but non-empty level.To achieve accurate control,we propose to combine the feedforward controller and the feedback controller and design a hybrid control-based algorithm HCA.Experiment results show that HCA is able to provide high bitrate and keep a very low rebuffering ratio within the low delay constraint.更多还原