A Reinforcement Learning Approach to Traffic Scheduling in Complex Data Center Topologies

Abstract

In this study, an optimization algorithm for traffic scheduling in data centers based on an actor-critic structure is proposed to improve network resource utilization and optimize task scheduling efficiency. Through experiments on the Google Cluster Usage Traces dataset, we analyze the adaptability and optimization effect of the algorithm under different network topology sizes, high concurrent task loads, and reward weight settings. Experimental results show that Actor-Critic can effectively improve throughput, reduce task loss rate, and show strong adaptability in the dynamic traffic environment. Compared with traditional scheduling methods, the algorithm has better scheduling stability in high-concurrency environments, but there is still room for optimization in very large-scale network environments. Future research can combine multi-agent reinforcement learning, federated learning, and graph neural networks to further improve the generalization ability of scheduling strategies, and provide more efficient solutions for intelligent data center management.