Via: Improving Internet Telephony Call Quality Using Predictive Relay Selection

The use of the Internet for voice calls is here to stay. In spite of the volume and importance of Internet telephony, we have little understanding of (1) how network performance impacts user-perceived call quality, and (2) why and where such quality problems occur in the wild. To bridge this gap, we analyze a data set of 4 3 0 million calls from Skype, with clients across 1 9 0 0 ASes and 1 2 6 countries. We observe that call quality problems are quite pervasive. More importantly, these problems are significantly spread out geographically and over time, thereby making simple fixes targeted at specific "pockets" of poor performance largely ineffective.To alleviate call quality problems, we present an architecture called VIA that revisits the use of classical overlay techniques to relay calls. We argue that this approach is both timely and pragmatic given the emergence of private backbones in recent years to connect globally distributed datacenters, which can serve as a readily available infrastructure for a managed overlay network. Trace-driven analysis shows that an oracle-based overlay can potentially improve up to 5 3 % of calls whose quality is impacted by poor network performance. A key challenge is realizing these benefits in practice, in the face of significant spatial and temporal variability in performance and a large number of relaying choices. We develop a practical relay selection approach that intelligently combines prediction-based filtering with an online exploration-exploitation strategy. Trace-driven analysis and a small-scale deployment shows that VIA cuts the incidence of poor network conditions for calls by 45 % (and for some countries and ASes by over 80 %) while staying within a budget for relaying traffic through the managed network.