Impact of buffering on quality of experience

Abstract

The Internet has become an essential part of the lives of millions of people and an invaluable asset to businesses. As an emerging trend, data storage and processing is shifting to the Cloud (e.g., Google Apps, or Cloud gaming), making users more and more dependent on the network to perform their daily activities. Despite the crucial importance of Internet services, they remain susceptible to bad service quality. One particular factor influencing service quality is buffering at various layers. This thesis assess the impact of buffering on Quality of Experience (QoE). QoE is an active research area aiming to quantify the users’ perception of Internet services. This is challenging since the users’ perception is subjective. This thesis tackles this challenge by using a multi-disciplinary approach that combines QoE and networking research to take a cross-layer perspective on network and application buffering. Network buffering occurs in hosts, switches, and routers throughout the Internet. It impacts network performance by contributing delays, jitter, and packet losses. Lossbased degradations of video quality are illustrated in a first evaluation. Motivated by this observation, Scalable Video Coding is discussed to optimize video QoE in phases of congestion. An evaluation of SVC dimensions shows that spatial scalability yields better QoE scores than temporal scalability. Further, QoE impacts of model based packet loss generators—e.g., as used in QoE studies—are evaluated. It is shown that the model choice impacts quality indicators, thus model choice matters. The size of network buffers influences network performance by controlling the level of introduced delay, jitter, and packet loss. The choice of ‘proper’ buffer sizing guidelines remains an unresolved and controversially discussed topic since decades. In this context, this thesis presents the first comprehensive study on the impact of buffer sizes on Quality of Experience, involving relevant user applications (e.g., voice, video, and web browsing), real hardware, and realistic workload. While bloated buffers can degrade QoE, buffer sizes that follow standard sizing guidelines significantly impact QoS metrics, but impact QoE metrics only marginally. Limiting congestion, may thus thus yield more immediate QoE improvements than optimal buffer sizes. Application buffering is used to compensate for performance variations, e.g., originating from network buffering. One example is the buffer-based, proprietary retransmission scheme in a major IPTV system. This thesis provides insights into the functioning of this scheme and motivates the extension of QoE metrics to account for client-side error recovery to prevent QoE mispredictions. To optimize Web QoE, a hit rate analysis of caching schemes is performed by focusing on YouTube video popularities. The thesis contributes an optimized caching scheme that offers higher cache hit rates than traditional Least Recently Used caches. Finally, it broadens the view of QoE by discussing spam as major QoE determinant in e-mail. A large-scale study conducted over 3.5 years reveals insights into address harvesting as the origin of spam and proposes mechanisms for spam mitigation that can help to improve e-mail QoE.