Understanding and optimizing Internet video delivery with self-driving networks

Video dominates today’s Internet and will play an even larger role going forward. Netflix users stream over 150 million hours of video per day, and video makes up more than half of the traffic on broadband networks. Networks struggle with handling this flood of traffic, wanting to simultaneously deliver high quality video to their users without letting it overwhelm other applications that require low latency service. Complicating matters, the network operators are stuck with outdated tools, and video is increasingly encrypted, giving little insight into which streams are behaving well and which might need help. In this project, we will develop techniques (based on machine learning and optimization methods) and conduct studies to understand the streaming video delivery ecosystem, addressing questions including: what strategies do services use to deliver video? Do simultaneous streams by different users (possibly using different streaming services) fairly share bandwidth, or do some crowd out others? How can operators infer the quality of experience they are providing to users when traffic is encrypted? Given that encrypted traffic, how can they design networks and traffic management techniques that automatically optimize video delivery, balancing bandwidth allocations to provide the best service possible to all users?