Gazing Beyond Horizon: The Predict Active Queue Management for Controlling Load Transients

Active Queue Management (AQM) is used by routers to limit queue build-up by reacting pro-actively to incipient congestion and modern AQM algorithms attempt to keep queuing latency at a low level. However, current AQM algorithms have hard time to correctly manage load transients, resulting in delay spikes. They suffer from a horizon problem that prevents a router from acquiring a complete picture of the traffic load by simply measuring its current queuing, because the distribution of the load over the end-to-end path keeps some of the load inducing packets out of the view of the router. Inability to know the RTTs of the flows going through the router makes it also challenging for the AQM algorithm to use a proper measurement interval for load calculation as it does not know what that interval should be. In this paper we tackle the horizon problem and RTT uncertainty that together complicate accurate load estimation for AQM algorithms. We propose a new AQM algorithm that not only determines the current load but also predicts the load into the near future. The prediction enables our algorithm to respond timely even to rapid load transients such as those due to TCP Slow Start. Our simulation results show that our algorithm achieves low queuing, whereas PIE and CoDelbased AQM algorithms struggle to limit queuing with small RTTs and with large RTTs react too early lengthening flow completion times.