Transformer-Enabled DRL Routing for DCNs with Sequential Flow Features

Routing is a key to achieving high-performance data center networks. However, due to the significant variations in network temporal information, traditional routing protocols fail to effectively capture the sequential characteristics and patterns of network traffic. To address this issue, a new dynamic routing approach is proposed through deep reinforcement learning with a modified transformer architecture, enabling an effective identification of sequential features in network traffic. By testing Software Defined Networking (SDN) in two simulated Data Center Network (DCN) topologies, we compare our approach with a traditional routing protocol and a nontemporal DRL routing protocol, demonstrating the efficacy and merits of our proposed solution.