Deep Reinforcement Learning Application for Network Latency Management

The centralization of network intelligence provided by decoupling the intelligence plane from the forwarding plane enabled by Software Defined Networking, and the recent breakthroughs of machine learning, paved the way to addressing a variety of network challenges. QoS-aware routing is a high complexity problem which is crucial for networks, due to the differences in the network requirements and the impact of the routing decisions on the whole network, therfore optimizing network performances (i.e., end-to-end delay, throughput) must be achieved in order to ensures QoS requirements. Neural Networks and Reinforcement learning are one of machine learning breakthroughs that can tackle this complexity problem. In this paper, we propose an efficient rules placement algorithm based on predictive network latency using Deep Reinforcement Learning (DRL) to cope the flow rules affectation problem and the Long Short-Term Memory (LSTM) for the estimation of the future traffic demands. To do so, we first formulate mathematically the QoS-aware routing where the optimization problem is to to minimize the total network delay, by efficiently allocate flows to different paths and proactively prevent congestion. Then, we propose a simple yet efficient heuristic algorithm to solve this optimization problem. The obtained results using ONOS controller and OpenVswitch revealed the efficiency of the proposed approach in decreasing network latency, packet loss and rising the network throughput.