Dependency-Aware Task Offloading in Multi-Controller IoV Networks

Abstract

In recent years, the Internet of Vehicles (IoV) has introduced a new class of latency-sensitive vehicular applications — including collision avoidance, cooperative driving, and real-time traffic analytics — that far exceed the computational capacity of on-board units (OBUs). Multi-access Edge Computing (MEC) has emerged as a promising paradigm for offloading such tasks to Road-Side Units (RSUs) deployed at the network edge, dramatically reducing the reliance on distant cloud servers. However, conventional offloading models treat tasks as indivisible units and assign them to a single MEC server, causing resource bottlenecks and deadline violations in dense, high-mobility vehicular environments. Fine-grained partial offloading addresses this limitation by decomposing tasks into sub-tasks, yet introduces further challenges: managing inter-sub-task dependencies, coordinating distributed execution across multiple edge nodes, and adapting decisions dynamically as vehicles traverse RSU coverage zones. Maintaining centralized control and a real-time global view of the network is essential to overcome these challenges, and a multi-controller Software-Defined Networking (SDN) architecture provides exactly such capabilities. This thesis proposes a mobility-aware SDN-based Partial Dependency-Aware Offloading (PDAO) framework for IoV edge networks. PDAO models vehicular applications as Directed Acyclic Graphs (DAGs) to capture sub-task dependencies, enabling fine-grained and dependency-consistent offloading decisions. Leveraging a distributed ONOS controller cluster for both vertical and horizontal offloading, PDAO dynamically schedules sub-tasks across multiple MEC servers while accounting for inter-sub-task communication costs, RSU load, and vehicle mobility. Extensive simulations demonstrate that PDAO significantly reduces task makespan, maintains a task completion ratio above 83\% under heavy load compared to standard single-controller SDN approaches, and achieves more balanced resource utilization compared to conventional baselines — underscoring the potential of multi-controller, dependency-aware partial offloading to meet the latency demands of safety-critical vehicular applications.