Inferring Netflix User Experience from Broadband Network Measurement

Abstract

Netflix is the largest video-streaming provider in the world today, with over 148 million subscribers and accounting for over 20% of broadband traffic in most developed countries. Internet Service Providers (ISPs) are acutely aware of the need to provide good video streaming experience to viewers, but are poorly equipped to measure and monitor per-stream quality. In this paper, we measure and analyze Netflix playback data from multiple households, develop a practical and scalable method to correlate network activity with client playback behavior, and provide a means for ISPs to infer per-stream Netflix experience from broadband traffic patterns. Our specific contributions are: (1) We develop a measurement tool for collecting network flow activity and client playback metrics, deploy it in 9 households and our lab to gather data for about 8000 Netflix video streams under various network conditions, and release the data to the public; (2) We analyze our data to highlight correlation between active flows and video playback phase, and between network chunk transfers and playback buffer health, during both regular-play and trick-play of video; (3) We develop a method for the ISP to infer Netflix user experience in terms of buffer fill-time, video bitrate and throughput, and detect playback buffer depletion and quality degradation events. ISPs can use our methods to measure, monitor, and manage Netflix user experience in real-time.