A congestion control algorithm based on deep reinforcement learning in SDN data center networks

Aiming at the TCP Incast problem in Software Defined Networks (SDN) data center networks, a new congestion control algorithm based on deep reinforcement learning is proposed in this paper. By the way of continuously interacting between the agent and the environment to obtain the optimal strategy, the SDN controller plays the role as agent and many-to-one model of data center networks plays the role as environment. Under the condition that queue length of the buffer of the bottleneck switch does not exceed the preset congestion threshold, data transmission rates of multiple servers are trained using deep reinforcement learning algorithms, and finally the optimal transmission rates which will not cause congestion are acquired. The simulation results show that the algorithm proposed in this paper can effectively avoid congestion in many-to-one scenarios in SDN data center networks and can improve the overall performance of networks.