CNRS LIP6, University Pierre et Marie Curie
Nov. 13, 11:00AM CS Open Meeting Area (CSB 477)
Abstract: Have you ever had your calls dropped on Skype? Have you ever had a video on YouTube freeze up on you? Network disruptions and dynamic network conditions can adversely impact end-user experience. These issues can frustrate users who are oblivious of the underlying causes, but have to deal with the resulting degradations. In an ideal world, user devices would have the capability to automatically detect and troubleshoot network performance problems or and provide contextual feedback to users when fixes are not available. There is much interest recently in doing automated diagnosis on user laptops and desktops. One interesting aspect of performance diagnosis that has received little attention is the user perspective on performance.
In this talk, we present HostView, a data measuring tool that collects network traffic headers and related information at end-hosts. Importantly, Hostview includes mechanisms for users to rate their perceived network conditions, and is a key departure from previous work in the area. We discuss the design tradeoffs in building HostView (overhead, privacy and user annoyance) and the challenges in collecting such data. Then, we present our efforts in developing predictors of user dissatisfaction with Internet application performance. We train these predictors using network performance data annotated with user feedback collected with HostView from the machines of 19 users. The main challenges of modeling user dissatisfaction with network performance comes from the scarcity of user feedback and the fact that poor performance episodes are rare. We develop a methodology to build training sets in face of these challenges. Then, we show that predictors based on non-linear support vector machine achieve higher true positive rates than predictors based on linear models. Our predictors consistently achieve true positive rates above 0.9. Finally we quantify the benefits of building per-application predictors over building general predictors that try to anticipate user dissatisfaction across multiple applications.
Speaker Biography: Renata is a CNRS researcher at LIP6, the computer science department of the University Pierre et Marie Curie. She finished her Ph.D. at the University of California San Diego in August 2005, where she was advised by professor Geoff Voelker. During her PhD, she split her time between San Diego and New Jersey (AT&T Labs - Research), where she worked with Jennifer Rexford and Tim Griffin on analyzing the impact of intradomain routing on BGP. Before being seduced by AT&T's network data, she was working with Geoff Voelker, Keith Marzullo and Stefan Savage on characterizing path diversity in IP networks.