Using Machine Learning to Control Congestion in SDN: A Review

Congestion is a major issue in networks, as it decreases efficiency and wastes bandwidth. While the basic operations of TCP remain the same, there are different flavors of TCP developed for specific network environments that help control congestion by updating window size and data transmission. Software-defined networking (SDN) can provide centralized control for network traffic, and the amount of data received by SDN controllers is huge. This huge data can be used as an input dataset for machine learning algorithms to extract a lot of information that helps improve network performance. To process this data, machine learning (ML) has been suggested to improve network performance and reinforcement learning (RL) to predict congestion. This article reviews recent ML algorithms for congestion control in SDNs, starting with a brief overview of SDN, ML, and congestion control and then reviewing the recent works that apply ML to control congestion. Based on this comprehensive review, it has been concluded that the RL actor-critic algorithm is the most efficient approach to prevent congestion in SDN networks. In addition, the ML random forest algorithm has successfully classified flow types and detected the flow that may cause congestion.