Web Quality of Experience Measurement: Metrics, Methods and Tools

Abstract

The web is one of the dominant applications on the Internet. Over the last three decades, the web has been evolving in terms of content types, supporting technologies, content provisioning, and access protocols. Similarly, the users' demands for fast and reliable web access have been also growing. Understanding the user browsing Quality of Experience (QoE) is of interest to content and service providers to deliver a quality service. However, the subjective nature of QoE makes it challenging to measure the web user experience on a large scale. Due to this, Quality of Service (QoS) metrics that can be measured on different layers of the web stack have been used to approximate the user experience. In this thesis, we propose a method to calculate an objective web QoE metric that better approximates the user experience. We design and implement a measurement system and tool that can be used on a large scale. We discuss the validation of the measurement system and benchmark the system performance. We present results from measurements that have been conducted to understand the web performance and QoE both from fixed-line and cellular networks. We also discuss modeling the web QoE from the QoS metrics using existing export models (e.g., ITU-T and IQX), and machine learning algorithms (e.g., SVR, CART, BOOST). This thesis contributes to the effort towards understanding, designing, and managing infrastructure to provide improved web QoE. Web users and content and service providers can use the methodology we have proposed and the tools we have designed to understand and troubleshoot possible bottlenecks for poor user experience. For instance, Internet Service Providers (ISPs) can deploy our tools on customer premises in their subscriber base and monitor their end-user web QoE. ISPs can use this for efficient capacity planning, network design, and web traffic management towards popular Content Delivery Networks (CDNs). The work on modeling web QoE shows that the expert models and machine learning-based models have comparable degree of performance accuracy. This thesis also shows that the expert models can accommodate new time-related metrics beyond the web latency metrics.