Towards resource-efficient and QoS-aware video adaptation in media cloud
Date of Issue2017-12-08
Interdisciplinary Graduate School (IGS)
Video streaming dominates Internet traffic, accounting for more than 70 percent of North American downstream traffic at peak time. However, limited bandwidth capacity, unstable network condition, and diverse viewing devices inherently deteriorate user experiences, triggering a tussle between the growing demand of video traffic and the quality of viewing experiences. Video adaptation is the de facto solution for video streaming over heterogeneous viewing devices and under time-varying network connections. For video adaptation, each video must be transcoded into multiple representations in different bitrates and resolutions. The client-side can dynamically select the best possible quality representation according to the current network condition and device capacity. Nevertheless, embracing video adaptation in video streaming faces many challenges regarding operational cost, Quality of Service (QoS), and Quality of Experience (QoE). First, video transcoding is compute-intensive, and transcoding source videos into multiple representations and storing them consume tremendous resources. Adopting video adaptation mechanism can thus greatly increase the operational cost for video streaming. To reduce the operational cost, we propose the partial transcoding scheme for cost-efficient video transcoding. Specifically, the frequently requested video chunks are cached, resulting in storage cost; while the seldom requested video chunks are transcoded online when being requested, resulting in computing cost. We aim to minimize the long-term overall cost by determining whether a video chunk should be cached or transcoded online. We design an online algorithm by leveraging the Lyapunov optimization framework to make the caching decision. We also design the virtual caching scheme, vCache, by considering the practical implementation under the Network Functions Virtualization (NFV) infrastructure. vCache can dynamically provision computing resources to ensure that transcoding delays will not affect streaming services. Second, the video generation rate in an online video service is time-varying, and maintaining a fixed number of servers for transcoding to meet the peak workload may waste tremendous resources. A new trend for transcoding is to adopt the cloud infrastructure for elastic resource provisioning and parallel transcoding. Thus, intelligent strategies are required to provision the right amount of resources to meet QoS requirements. We study the resource provisioning problem for transcoding in three scenarios. We first propose a two-timescale optimization framework for maximizing profit of transcoding service while meeting QoS requirements by jointly provisioning resources and scheduling tasks. This method analytically integrates service revenue, processing delay, and resource consumption into one optimization framework. We then leverage the Model Predictive Control (MPC) to design an online algorithm for dynamic resource provisioning using prediction to accommodate to time-varying workloads. We improve our online algorithm through robust design to seek robustness of system performance against prediction noise. Finally, we develop a framework for scheduling the transcoding for live content and Video-on-Demand (VoD) content with statistical QoS guarantees. Each type of videos is specified with a QoS criterion and a QoS loss bound. Our method can provision the minimum amount of resources while keeping QoS loss probabilities within the prescribed bounds. Third, traditional rate adaptation approaches are semantics-agnostic, treating videos as common data. However, a viewer may have different degrees of interest on different parts of a video due to different video semantics. The interesting parts of a video can draw more visual attention of a viewer, and thus have higher visual importance. As such, delivering a viewer's interested parts of a video in a higher quality can improve the perceptual video quality compared with the semantics-agnostic approaches which treat each part of a video equally. We propose an interest-aware rate adaptation approach for improving QoE by inferring viewer interest based on video semantics. We first use the deep learning method to recognize video scenes, followed by leveraging the Term Frequency-Inverse Document Frequency (TF-IDF) method to analyze the degrees of an individual viewer's interest on different video scenes. The bandwidth, buffer occupancy, and viewer interest information are jointly considered under the MPC framework for selecting appropriate bitrates for maximizing QoE. We implement our video transcoding and streaming system, and conduct extensive experiments to evaluate the performances of our proposed methods. The experiment results show that our proposed methods can reduce the operational cost and guarantee QoS for video transcoding, and improve QoE for adaptive video streaming.