AI-Based Dynamic Load Balancing in Cloud Data Centers

Abstract

Cloud data centers underpin the digital infrastructure of contemporary organizations, hosting diverse applications ranging from e-commerce platforms to large-scale scientific computations. With fluctuating and often bursty request patterns, static or heuristic load balancing schemes struggle to maintain optimal resource utilization, low latency, and SLA compliance. This paper introduces an AI-driven dynamic load balancing framework based on a Deep Q-Network (DQN) agent that continuously observes multi-dimensional host states—CPU utilization, memory occupancy, queue lengths, and network I/O—and makes real-time decisions on request routing and VM migration. We implement the framework in CloudSim 5.0, modeling a mid-sized data center of 20 hosts and 200 VMs under both Poisson and heavy-tailed arrival distributions.

Over 30 independent trials, our approach reduces average response time by 29% relative to Round Robin and 23% relative to Least Connections, boosts throughput by up to 33%, and raises CPU utilization from ~70% to ~83%. Rigorous statistical validation (one-way ANOVA, Tukey’s HSD, effect-size analysis) confirms the significance and robustness of these gains. In addition to performance improvements, the DQN agent exhibits rapid adaptation to workload surges, converging to stable policies within 600 training episodes. This study demonstrates that reinforcement learning can serve as a viable, generalizable strategy for real-time, multi-objective resource management in cloud environments, paving the way for self-optimizing data centers.