A Load-Aware Replica Selection Strategy with Multi-Armed Bandits and Adaptive Redundancy in ICN

With the rapid growth of data-intensive applications, ensuring low-latency and reliable content retrieval in complex networks is a critical challenge. Information-Centric Networking (ICN) leverages content naming and pervasive innetwork caching to enable retrieval from multiple replicas, making replica selection a key determinant of performance. However, effective selection remains difficult, as replica nodes are constrained by limited capacity, bursty workloads, and dynamic path variations. Traditional approaches based on historical measurements or prediction require frequent probing, which incurshigh overhead while still failing to explore new replicas effectively. To mitigate this, the Multi-Armed Bandit (MAB) framework has been adopted to balance exploration and exploitation. Yet, existing MAB-based methods still face high exploration costs in large-scale scenarios and lack explicit control of tail latency. To address these challenges, we propose a load-aware replica selection strategy that integrates the MAB framework with adaptive redundancy control. The MAB framework enables online learning of replica dynamics, allowing adaptation to timevarying conditions without frequent global probing. Redundancy is adjusted according to system load, being increased under light load to suppress tail latency and reduced under heavy load to prevent congestion. Simulations on a real-world topology demonstrate that the proposed strategy significantly lowers both mean and tail latency, improves load balancing, and enhances robustness under diverse conditions.