SLA-Aware RSU-Edge Delegate Orchestration for IoV Consensus

Abstract

Ensuring reliable and timely consensus among Internet of Vehicles (IoV) nodes is critical for safety and operational efficiency, particularly under high mobility and dynamic network conditions. Traditional consensus protocols, however, do not explicitly incorporate service-level objectives (SLOs) such as commit latency, tail latency, or delegate-set diversity, limiting their applicability in real-world deployments. In this paper, we present a service-level agreement (SLA)-aware road-side unit (RSU)-edge orchestration framework for IoV consensus delegate selection, which leverages reinforcement learning (RL) to optimize committee composition while preserving quorum safety. Our approach embeds SLO metrics directly into the proximal policy optimization (PPO) reward function, enabling the RSU-edge to adapt delegate selection online under varying vehicle densities, speeds, and network conditions. A shortlist-based candidate reduction mechanism reduces computational overhead, while certificate-governed reconfiguration and state transfer support safe committee activation and recovery. Extensive simulations across multiple scenarios, including burst losses and mobility-induced churn, demonstrate that our method reduces median and tail commit latency, increases throughput, and maintains higher delegate-set diversity than baseline heuristics and the adapted BFTBrain-style service comparator. Under the simulator reference configuration, the modeled proposal-construction components yield a component-wise tail budget of 18.8 ms, excluding governance certification and state synchronization. The framework provides a practical blueprint for service-level-aware management of IoV consensus, bridging the gap between protocol-level designs and operational network management. Within the controlled service-level simulation scope, the study demonstrates the feasibility, robustness, and performance advantages of RL-driven RSU-edge orchestration. Packet-level and field deployment validation remain future work.