The use of the public 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 voice call quality, and (2) why and where such quality problems occur in the wild. To bridge this gap, we analyze a data set of 88 million calls from a large VoIP provider with clients spread across 1900 ASes and 126 countries. One of the key findings is that calls over bad networks are spread out geographically and over time.
To alleviate call quality problems, we present an architecture called VIA that revisits the use of classical overlay techniques. We argue that this approach is both timely and pragmatic given that the private backbones built up in the recent years to connect globally distributed datacenters provide a readily available infrastructure for a managed overlay network. Trace-driven analysis shows that V IA can potentially improve up to 83% of calls whose quality is impacted by poor performance on key network metrics. A key research question, however, is whether these benefits can actually be realized in practice, in the face of significant spatial and temporal variability in performance and the existence of a large number of relaying choices. To this end, we develop a practical relay selection approach that intelligently combines prediction-based filtering with an online explorationexploitation strategy. We demonstrate using both a largescale trace-driven analysis and a small-scale real-world pilot deployment that V IA helps cut the incidence of poor network conditions for calls by up to 66% (and for some countries and ASes by even over 80%) while staying within a budget for relaying traffic through the managed infrastructure.