Title: Reinforced Cooperative Load Balancing in Data Center

Abstract: Network load balancers are central components in modern data centers, that cooperatively distribute workloads of high arrival rates across application servers, thereby contribute to offering scalable services. The independent and "selfish" load balancing strategy is not necessarily the globally optimal one. This paper represents the load balancing problem as a cooperative team-game with limited observations over system states, and adopts multi-agent reinforcement learning methods to make fair load balancing decisions without inducing additional processing latency. On both a simulation and an emulation system, the proposed method is evaluated against other load balancing algorithms, including state-of-the-art heuristics and learning-based strategies. Experiments under different settings and complexities show the advantageous performance of the proposed method.