Section: Application Domains

Quality of Experience

An increasing number of residential users consume online services (e.g., VoD, Web browsing, or Skype) in their everyday activities (e.g., for education or entertainment purposes), using a variety of devices (e.g., tablets, smartphones, laptops). A high Quality of Service (QoS) is essential for sustaining the revenue of service providers, carriers, and device manufactures. Yet, the perceived Quality of Experience (QoE) of users is far from perfect e.g., videos that get stalled or that take a long time to load. Dissatisfied users may change Internet Service Providers (ISPs) or the online services. Hence, the incentives for measuring and improving QoE in home networks are high while mapping network and application QoS to QoE is a challenging problem. In this work we have focused in measuring several network Quality-of-Service (QoS) metrics, such as latency and bandwidth, both in residential Wi-Fi as well as broadband networks, homes are using for connecting to the Internet.

The WiFi Context. Residential Wi-Fi performance, however, is highly variable. Competing Wi-Fi networks can cause contention and interference while poor channel conditions between the station and the access point (AP) can cause frame losses and low bandwidth. In some cases, the home Wi-Fi network can bottleneck Internet access. While problems in the Wi-Fi network may affect several network QoS metrics, users will typically only notice a problem when poor Wi-Fi affects the QoE of Internet applications. For example, a Wi-Fi network with low bandwidth may go unnoticed unless the time to load Web pages increases significantly. A user observing degraded QoE due to Wi-Fi problems may mistakenly assume there is a problem with the Internet Service Provider (ISP) network. Our discussions with residential ISPs confirm that often customers call to complain about problems in the home Wi-Fi and not the ISP network.

Prior work has focused on QoS metrics for some applications (e.g., on-line video, Web browsing, or Skype) with no attempt to identify when Wi-Fi quality affects QoE. We are particularly interested in assisting ISPs to predict when home Wi-Fi quality degrades QoE. ISPs can use this system to detect customers experiencing poor QoE to proactively trigger Wi-Fi troubleshooting. ISPs often control the home AP, so we leverage Wi-Fi metrics that are available on commercial APs. Detecting when Wi-Fi quality degrades QoE using these metrics is challenging. First, we have no information about the applications customers are running at any given time. ISPs avoid capturing per-packet traffic traces from customers, because of privacy considerations and the overload of per-packet capture. Thus, we must estimate the effect of Wi-Fi quality on QoE of popular applications, which most customers are likely to run. In this context, we study Web as a proof of concept, as a large fraction of home traffic corresponds to Web. Second, application QoE may be degraded by factors other than the Wi-Fi quality (e.g., poor Internet performance or an overloaded server). Although a general system to explain any QoE degradation would be extremely helpful, our monitoring at the AP prevents us from having the end-to-end view necessary for such general task. Instead, we focus on identifying when Wi-Fi quality degrades QoE. Finally, Wi-Fi metrics available in APs are coarse aggregates such as the average PHY rate or the fraction of busy times. It is open how to effectively map these coarse metrics into QoE.

Predicting QoE. Clearly, different actors in the online service chain (e.g., video streaming services, ISPs) have different incentives and means to measure and affect the user QoE. Uncovering statistically equivalent subsets of QoS metrics across and within levels provides actionable knowledge for building QoE predictors. To achieve this goal, we leverage recent advances on feature selection algorithms to exploit available experimental evidence of the joint probability distributions of QoE/QoS metrics. This type of statistical reasoning will enable us to determine local causal relationships between a target QoE variable, seen as effect, and multiple QoS metrics across or within levels, seen as causes. Such data-driven analysis is justified by the multiplicity of dependencies that exist between network or application QoS metrics as different adaptation mechanisms (e.g., TCP congestion avoidance, HTTP bitrate adaptation) are activated at each level in real life. Building optimal predictors based on (eventually several) probabilistically minimal subsets of features opens the way for a principled comparison of the predictors.