Topology Design for Data Center Networks Using Deep Reinforcement Learning

Abstract

This paper is concerned with the topology design of data center networks (DCNs) for low latency and fewer links using deep reinforcement learning (DRL). Starting from a K-vertex-connected graph, we propose an interactive framework with single-objective and multi-objective DRL agents to learn DCN topologies for given node traffic matrices by choosing link matrices to represent the states and actions as well as using the average shortest path length together with action penalty terms as reward feedback. Comparisons with commonly used DCN topologies are given to show the effectiveness and merits of our method. The results reveal that our learned topologies could achieve lower delay compared with common DCN topologies. Moreover, we believe that the method can be extended to other topology metrics, e.g., throughput, by simply modifying the reward functions.