ARS: Adaptive routing strategies using AT-GCN for traffic optimization in data center networks

Abstract

As internet services and Internet of Things (IoT) devices rapidly expand, Data Center Networks (DCN) have become essential for supporting online services, cloud computing, and big data analysis. These devices generate massive amounts of data continuously, leading to uneven and sudden network loads that traditional networks struggle to handle. Existing routing strategies often rely on fine-grained routing or current network states, which can result in misjudgments and fail to manage sudden traffic spikes or meet both low-latency and high-bandwidth requirements effectively. To tackle these challenges, we propose a new routing strategy that combines Software-Defined Networking (SDN) with an Attention-based Temporal Graph Convolutional Network (AT-GCN). By predicting future network states using AT-GCN and analyzing traffic characteristics with a Deep Neural Network (DNN), our method offered more precise traffic scheduling. Specifically, we used an improved butterfly optimization algorithm to route mouse flows through low-latency, stable paths, and employ Dijkstra’s algorithm to send elephant flows along paths with the highest predicted bandwidth. This approach effectively reduces latency and increases throughput. Simulation results demonstrate that our method significantly reduces average Flow Completion Time (FCT) by 18.72% to 62.94% compared to existing ECMP, LetFlow, DLBF, ILB and MABC schemes.