Dynamic Load Balancing in Software-Defined Networks Using Machine Learning

Abstract

In order to enhance the performance of data transmission in software-defined networks (SDN), we propose a solution for load balancing by leveraging the property of global view of network in SDN. The aim is to identify best path for data transmission and get significant decrease in network latency. For this, several relevant features from each path are collected first. The throughput, rate of packet loss, latency, the number of hops and node utilization are extracted by the load balancer. Using the features mentioned above, the overall load condition is predicted by the trained neural network model for various shortest paths provided by the Dijkstra’s algorithm. The Dijkstra’s algorithm uses transmission rate of the link as the metric. The transmission rate of each link extracted using API provided by Mininet [1]. Then, the artificial neural network is tested in real time to achieve maximum link utilization and node utilization. Maximum node utilization is achieved node with the least load is selected as the total node utilization along each path is also extracted and fed to the neural network. After selecting the path with minimum load, the load balancer sends the path information to the SDN controller. The SDN controller then goes on to push the flow rules in the switches along the best path given by the load balancer. If there is a link or node failure along the best path, then the load balancer selects the second-best path in real time so that the data transfer is completed with minimum delay. Thus, optimum overall network utilization is achieved by this system in real time.