LFQ: Online Learning of Per-flow Queuing Policies using Deep Reinforcement Learning

摘要: The increasing number of different, incompatible congestion control algorithms has led to an increased deployment fair queuing. Fair queuing isolates each network flow and can thus guarantee fairness for even if the flows’ controls are not inherently fair. So far, queue in system either a fixed, static maximum size or is managed by Active Queue Management (AQM) algorithm like CoDel. In this paper we design AQM mechanism (Learning Qdisc (LFQ)) that dynamically learns optimal buffer according specified reward function online. We show our Deep Learning based assign depending on its control, delay bandwidth. Comparing competing schedulers, it provides significantly smaller queues while achieving same higher throughput.