Video viewing is on an upswing, thanks in large part to the relative ubiquity of speedy, affordable internet connectivity. By 2021, a million minutes (17,000 hours) of video content will cross worldwide networks every second, according to Cisco. And it’s estimated that video streams accounted for 75% of all traffic in 2017, a share anticipated to rise to 82% by 2022.
In an effort to develop tech suited to delivering tens of thousands of petabytes of video each month, scientists at the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory (CSAIL) recently investigated a system that leverages video player state data and file characteristics to optimize congestion control. (In this context, “fairness” refers to how similar the viewing experience is for different users.) They report that their end-to-end protocol — Minerva — substantially cuts down on both buffering and pixelation without requiring changes to underlying infrastructure.
“The growth of video traffic makes it increasingly likely that multiple clients share a bottleneck link, giving video content providers an opportunity to optimize the experience of multiple users jointly,” wrote the researchers in a preprint paper. “But today’s transport protocols are oblivious to video streaming applications and provide only connection-level fairness.”
As the team explained further, most video content providers are beholden to bandwidth decisions made by congestion-reducing algorithms like Reno and Cubic, which seek to achieve connection-level fairness by giving competing flows an equal share of a link’s capacity. As a result, providers fine-tune viewing experiences in isolation, rather than allocating bandwidth among clients, and they don’t take into account factors like genre, screen size, screen resolution, and device type or playback buffer size.
By contrast, Minerva dynamically adjusts video streaming rates for fairness even without explicit information about competing video clients. When several of these clients share a bottleneck link, their rates converge to a bandwidth allocation that doesn’t interfere with other internet traffic.
Specifically, Minerva implements techniques and distributed algorithms that capture the relationship between bandwidth and quality of experience. Each client computes dynamic weights for its videos through the course of the videos, and it determines bandwidth allocations proportional to the weight from network conditions and other variables.
In experiments involving a real-world residential Wi-Fi network and two Amazon Web Services instances connected to eight clients, the researchers report that a quarter of the time Minerva improved quality for 15-32% of the videos by “an amount equivalent to a bump in resolution from 720p to 1080p.” Moreover, they say the protocol reduced total rebuffering time an average 47%, even with unpredictable data arrivals and departures, by allocating bandwidth to videos at risk of rebuffering,
“If five people in your house are all streaming video at once, [Minerva] can analyze how the various videos’ visuals are affected by download speed,” said MIT professor Mohammad Alizadeh, a senior author on a related paper that’s scheduled to be presented at the Association for Computing Machinery’s Special Interest Group on Data Communications (SIGCOMM) in Los Angeles later this month. “It then uses that information to provide each video with the best possible visual quality without degrading the experience for others.”