Edge AI-Assisted Active Queue Management

The core functions in the Internet have traditionally been designed to support high throughput. However, an increasing number of Internet services such as Web or Web-based applications and most interactive applications are sensitive to network latency.

In this study, initial steps towards taking a revolutionary approach to techniques that provide low latency for Internet traffic were planned and investigated. In contrast to pre-existing techniques, this new approach takes into account transient network behavior such as new connections that rapidly accelerate their transfer rate. Because such transients are very common in the Internet, any technique wishing to provide low latency should address them.

If transient behavior is not considered properly, arising latency spikes may cause interference such as short pauses and other disruptions for any co-existing latency-sensitive services (e.g., audio/video conferencing) or general “slowness” of transmission that notably impact the quality of service and annoy the Internet user.

However, if low latency can be provided under all circumstances, not only the quality of existing services improves but also the Internet-wide deployment of many emerging applications (e.g., natural video-based interaction, video-based remote control, cloud-based AR/VR) and entirely new applications would become possible.

The approach taken in this study is build on a new active-queue management (AQM) algorithm for Internet routers called Predict that was recently proposed in the PhD thesis by Ilpo Järvinen. The Predict algorithm applies novel heuristics for predicting the network load into the near future, allowing the AQM to react timely to avoid disruptive network delay spikes. The planned extensions include AI-based assistance for improved reaction to more wider set of network traffic patterns.

An initial implementation of some of the planned extensions were experimented with promising preliminary results. In the experiments, the peak queuing delay that is typically a significant component of the total latency was notably lower with the enhanced Predict algorithm compared to the exixting state-of-the-art AQM algorithms.

More information:
Markku Kojo
Supervisor for Doctoral Programme in Computer Science