Improving video streaming experience through network measurements and analysis

Multimedia traffic dominates today’s Internet. In particular, the most prevalent traffic carried over wired and wireless networks is video. Most popular streaming providers (e.g. Netflix, Youtube) utilise HTTP adaptive streaming (HAS) for video content delivery to end-users. The power of HAS lies in the ability to change video quality in real time depending on the current state of the network (i.e. available network resources). The main goal of HAS algorithms is to maximise video quality while minimising re-buffering events and switching between different qualities. However, these requirements are opposite in nature, so striking a perfect blend is challenging, as there is no single widely accepted metric that captures user experience based on the aforementioned requirements. In recent years, researchers have put a lot of effort into designing subjectively validated metrics that can be used to map quality, re-buffering and switching behaviour of HAS players to the overall user experience (i.e. video QoE). This thesis demonstrates how data analysis can contribute in improving video QoE. One of the main characteristics of mobile networks is frequent throughput fluctuations. There are various underlying factors that contribute to this behaviour, including rapid changes in the radio channel conditions, system load and interaction between feedback loops at the different time scales. These fluctuations highlight the challenge to achieve a high video user experience. In this thesis, we tackle this issue by exploring the possibility of throughput prediction in cellular networks. The need for better throughput prediction comes from data-based evidence that standard throughput estimation techniques (e.g. exponential moving average) exhibit low prediction accuracy. Cellular networks deploy opportunistic exponential scheduling algorithms (i.e. proportional-fair) for resource allocation among mobile users/devices. These algorithms take into account a user’s physical layer information together with throughput demand. While the algorithm itself is proprietary to the manufacturer, physical layer and throughput information are exchanged between devices and base stations. Availability of this information allows for a data-driven approach for throughput prediction. This thesis utilises a machine-learning approach to predict available throughput based on measurements in the near past. As a result, a prediction accuracy with an error less than 15% in 90% of samples is achieved. Adding information from other devices served by the same base station (network-based information) further improves accuracy while lessening the need for a large history (i.e. how far to look into the past). Finally, the throughput prediction technique is incorporated to state-of-the-art HAS algorithms. The approach is validated in a commercial cellular network and on a stock mobile device. As a result, better throughput prediction helps in improving user experience up to 33%, while minimising re-buffering events by up to 85%. In contrast to wireless networks, channel characteristics of the wired medium are more stable, resulting in less prominent throughput variations. However, all traffic traverses through network queues (i.e. a router or switch), unlike in cellular networks where each user gets a dedicated queue at the base station. Furthermore, network operators usually deploy a simple first-in-first-out queuing discipline at queues. As a result, traffic can experience excessive delays due to the large queue sizes, usually deployed in order to minimise packet loss and maximise throughput. This effect, also known as bufferbloat, negatively impacts delay-sensitive applications, such as web browsing and voice. While there exist guidelines for modelling queue size, there is no work analysing its impact on video streaming traffic generated by multiple users. To answer this question, the performance of multiple videos clients sharing a bottleneck link is analysed. Moreover, the analysis is extended to a realistic case including heterogeneous round-trip-time (RTT) and traffic (i.e. web browsing). Based on experimental results, a simple two queue discipline is proposed for scheduling heterogeneous traffic by taking into account application characteristics. As a result, compared to the state-of-the-art Active Queue Management (AQM) discipline, Controlled Delay Management (CoDel), the proposed discipline decreases median Page Loading Time (PLT) of web traffic by up to 80% compared to CoDel, with no significant negative impact on video QoE