Dynamic buffer sizing using Reinforcement Learning in ORAN

The 5th Generation (5G) cellular networks offer ultra reliable low-latency communication (URLLC) for emerging latency-sensitive applications. Various techniques within 5G access network (5G-AN) protocol stack have been deployed to guarantee the quality of service (QoS) of the access link. However, end-to-end (E2E) latency can be significantly high in the cellular scenario due to the unnecessarily large buffer size, called bufferbloat problem. Despite a large buffer in 5G-AN can guarantee the sufficient utilisation of physical resources, it introduces high queueing delay and misleads the source of a TCP flow to overestimate the available capacity. Reinforcement learning (RL) has the potential to recommend a dynamic buffer size suitable for the given channel state condition. Open radio access network (O-RAN) provides a flexible platform for reconfiguration to 5G-AN stack with new interfaces, that allows both non-real time and near-real time interaction between various elements within the AN based on actual channel conditions. Our scheme termed dynamic buffer sizing using RL (DBS-RL) is designed to operate coordinating with 5G-AN stack to adjust buffer size dynamically in a cellular scenario. DBS-RL achieved higher throughput and comparable delay compared with RED and CoDel without fine-tuned parameters. DBS-RL also outperforms static oversized buffer configurations.