Real-Time Analytics of Concurrent Adaptive Video Streams using Next-Generation Protocols

My master's thesis: an analysis service that ingests the logs of video streaming clients in real time and visualizes them.

Master’s thesis for the Master of Science in Computer Science (Networking and Security) at Hasselt University, academic year 2021-2022. Promotor: Peter Quax. Co-promotor: Wim Lamotte. Mentors: Joris Herbots and Maarten Wijnants.

Thesis (UHasselt)

Summary

Streaming video over HTTP/3 and QUIC can remove flaws of the traditional HTTP/1.1 over TCP, but the flaws have to be found first. Streaming clients produce a lot of logging data that can be analyzed for this.

The thesis proposes an analysis service that ingests this logging data in real time and visualizes it. It is linked to the testing framework Vegvisir, so that the logs of video streaming applications can be inspected. This makes it possible to follow the sessions of several streaming clients while they stream, and to infer events caused by outside forces such as network middleboxes, other streaming clients and applications competing for bandwidth.

How it works

The analysis service is built on the Vegvisir testing framework and has two parts. The analyzer receives the logs of the streaming clients and, optionally, of the server and the network. Clients send them in the qlog format, which many QUIC implementations support, and the thesis extends the ABR qlog schema with new events, such as the decisions of the ABR algorithm. The analyzer exposes the data through a GraphQL endpoint. The visualizer is a set of Grafana dashboards that query it, so graphs update while a test runs.

Streaming clients, a network simulator and a server form the test setup; their logs go to the analyzer, which the visualizer queries to show dashboards to an analyst

Real-time analysis works without hurting the streaming, as long as the network has the bandwidth to carry the logs. Comparing several clients, or one client under different settings, gave more insight than following a single live stream. The thesis ends by proposing machine-learning anomaly detection to point an analyst to the interesting parts of the data.

A Grafana dashboard that shows the QUIC and ABR qlog events of one streaming session side by side

The combination dashboard: the QUIC and ABR traces of the same streaming session, side by side.

A Grafana dashboard of QUIC qlog events, with round-trip times, events and data in flight

One of the first dashboards: the QUIC events of a stream, with round-trip times and data in flight.

Cite

@mastersthesis{vandersanden2022realtime,
  author = {Vandersanden, Mike},
  title  = {Real-Time Analytics of Concurrent Adaptive Video Streams using Next-Generation Protocols},
  school = {Hasselt University},
  year   = {2022},
  url    = {http://hdl.handle.net/1942/38549},
}
BibTeX