National Science Foundation
Consumer video dominates Internet traffic and shows no signs of abating. It is predicted that video will represent upwards of 70% of the Internet's traffic by 2021. Consumer video is primarily from video-on-demand (VoD) services but also includes a small but growing proportion of live video. The video for these services is streamed using techniques that comprise HTTP client-server interactions and that adapt the video bit rate to network conditions. There are a number of variations of such techniques which we will refer to collectively as HTTP Adaptive Streaming (HAS).
HAS schemes enable video to be streamed on networks that are shared among video streaming clients, as well as users of other network services, without special network provisioning or resource reservation to specifically accommodate the video traffic. When network conditions degrade, the adaptation schemes' primary aim is to keep the video stream “going” at the expense of the video quality. The adaptation schemes will also make use of improved network conditions to improve video quality. Video service quality is estimated through a video Quality of Experience (QoE) measure. Interpreted literally, video QoE is “a measure of the delight or annoyance” of a video customer's experience. Measuring video QoE is therefore difficult, as user experience is subjective and hard to quantify. Recent standardization efforts propose to use objective video QoE metrics such as video quality and video stalls to model user QoE.
Due to the flattening of the Internet, most consumer video traffic traverses only two networks, a video content provider network where the video source attaches and an Internet Service Provider (ISP) where the end user is connected (or end-user ISP for short) . The responsibility for delivering the best possible video QoE to users is shared between these two providers. In some instances the video is served from within the end-user ISP by a server provided by the video content provider. In this case it is mostly the end-user ISP's responsibility to insure the delivery of video QoE to users.
End-user ISPs need an in-depth understanding of end-user QoE and its relationship to network performance. This is a first step in the ISP's own network management in support of video streaming applications -- for short term trouble shooting and problem mitigation, and for long term QoE-aware network planning and provisioning. However, estimating video QoE is challenging for a network operator, since they typically do not have access to the streaming app on user devices, the device itself, or the server. Hence, unlike content providers, they cannot use in-app plug-ins for measuring the streaming QoE. Recent work proposes an alternative networking paradigm in which network operators and content providers collaborate, allowing content providers to share QoE information with network operators. However, this approach requires significant effort and is not immediately realizable. End-user ISPs are, therefore, currently constrained to use data derived from within their network to estimate videoQoE.
Our goal in the proposed work is to develop approaches that use network measurement to infer video QoE. In developing inference approaches we aim to provide a toolbox of validated and scalable techniques suitable for a diversity of network contexts as well as types of network measurement. Our work will consider two network environments, a Mobile Network Operator (MNO) network (through a collaboration with AT&T) and the Georgia Tech campus network. These two types of networks will allow us to investigate a wide range of user network connectivity including cellular, WiFi, and Ethernet. They will also allow us to consider different types of users on a variety of device types and usage modalities.
Our work will address three main challenges in the design and deployment of video QoE inference: