Download files
Abstract
Internet Service Providers (ISPs) continue to get complaints from users on poor experience for diverse Internet applications ranging from video streaming and gaming to social media and teleconferencing. Identifying and rectifying the root cause of these experience events requires the ISP to know more than just coarse-grained measures like link utilizations and packet losses. Application classification and experience measurement using traditional deep packet inspection (DPI) techniques is starting to fail with the increasing adoption of traffic encryption and is not cost-effective with the explosive growth in traffic rates. This thesis leverages the emerging paradigms of machine learning and programmable networks to design and develop systems that can deliver application-level intelligence to ISPs at scale, cost, and accuracy that has hitherto not been achieved before.
This thesis makes four new contributions. Our first contribution develops a novel transformer-based neural network model that classifies applications based on their traffic shape, agnostic to encryption. We show that this approach has over 97% f1-score for diverse application classes such as video streaming and gaming. Our second contribution builds and validates algorithmic and machine learning models to estimate user experience metrics for on-demand and live video streaming applications such as bitrate, resolution, buffer states, and stalls. For our third contribution, we analyse ten popular latency-sensitive online multiplayer games and develop data structures and algorithms to rapidly and accurately detect each game using automatically generated signatures. By combining this with active latency measurement and geolocation analysis of the game servers, we help ISPs determine better routing paths to reduce game latency. Our fourth and final contribution develops a prototype of a self-driving network that autonomously intervenes just-in-time to alleviate the suffering of applications that are being impacted by transient congestion. We design and build a complete system that extracts application-aware network telemetry from programmable switches and dynamically adapts the QoS policies to manage the bottleneck resources in an application-fair manner. We show that it outperforms known queue management techniques in various traffic scenarios. Taken together, our contributions allow ISPs to measure and tune their networks in an application-aware manner to offer their users the best possible experience.
