Abstract
This paper proposes a distributed extension of near-Real Time Radio Access Network (RAN) Intelligent Controller (near-RT RIC or RIC) in O-RAN, featuring fine-grained functional disaggregation of xApps across multiple RICs. To address scalability and latency, we design independent IP E2AP interfaces for each RIC, reducing queuing delays and decoupling processing timelines. Furthermore, we introduce a Deep Q-Network (DQN) based load balancing scheme that leverages Deep Reinforcement Learning to optimize packet routing between E2 agents and distributed RICs. Experimental results from a real-world deployment demonstrate that dedicating a RIC to a specific xApp significantly reduces average latency compared to a RIC managing a batch of xApps. These findings highlight the scalability and latency benefits of modular, functionally decomposed RIC architectures for time-critical O-RAN services. Furthermore, we develop a framework that models realistic O-RAN scenarios with configurable E2AP traffic, multiple RICs hosting hundreds of xApps. A DQN agent is trained with latency, load balancing, and congestion avoidance objectives to optimize packet routing under high-load and high-density xApp deployments. Compared to random and round-robin baselines, the DQN achieves up to 45% lower latency, and improved scalability across 5-100 RICs, demonstrating the effectiveness of Deep Reinforcement Learning for distributed RIC control.