A reinforcement learning-based network load balancing mechanism

Abstract

With the exponential growth of cloud computing demands, load balancing gradually becomes the infrastructure of high concurrency applications. The increased demand makes load balancing the cornerstone of large-scale systems' stability and efficiency. Due to the diversity of modern client-server architecture, irregular fluctuations in cluster system resource allocation gradually become significant. The existing load balancing algorithms are rule-based, leading to a gap between operational and practical scenarios. The phenomenon causes noticeable tribulations in cluster resource utilization optimization. This research proposes a reinforcement learning-based load balancing model optimization approach. This paper modelled the load balancing problem as a Markov Decision Process and implemented the conjecture with the Q-learning algorithm. This paper performs a load balancer that can efficiently utilize the resources in the cluster system by observing the association between the packet body and node resource utilization rate in the cluster system. The experiment demonstrated that the author's mechanism substantially improves the average cluster system resource utilization efficiency and reduces the Round-Trip Time performance in real network environments