Deep Reinforcement Learning with Graph Convolutional Networks for Load Balancing in SDN-Based Data Center Networks

Abstract

Since the 20th century, data center networks have been becoming more and more crucial with the data explosion in the information age. In order to manage large-scale data center networks, SDN-based data center networks is built. The larger scale of data transmitted in the network lead to the horizontal expansion of the network, which strengthen the capacity of network and make load balancing scenario being challenged. Load balancing algorithm can make the traffic evenly distributed in the network through routing flows through a suitable path. However, classical load balancing algorithms such as ECMP cannot adapt to the gradually expanding network scale. In recent years, the combination of artificial intelligence (AI) and load balancing has become a development direction, which have been proposed to make routing decisions intelligently according to the state of the dynamic network traffic. In this paper, we propose GDLB, a deep reinforcement learning (DRL) model for the load balancing problem of SDN data center networks, which combines graph convolutional neural networks (GCN) and DRL. Due to the ability of GCN to perceive network status and network topology, GDLB can achieve load balancing by making intelligent routing decisions. Compared with other AI-based algorithms, our algorithm can perceive network topology information and enable the model to make more superior decisions. Through using multiple indicators to evaluate rewards, our model can mainly pay attention to load balancing, but also ensures that QoS indicators such as delay and packet loss rate are qualified. Experiments and simulations show that our model has a greater improvement in maximum link utilization and QoS than other algorithms.