SDN-BASED DEPENDENCY-AWARE PARTIAL OFFLOADING IN IOT EDGE NETWORKS

Abstract

In recent years, Multi-access Edge Computing (MEC) has gained enormous attention due to its capability to manage latency-sensitive Internet of Things (IoT) applications such as automated vehicles, augmented reality, and virtual reality. These applications can be offloaded to MEC servers (deployed at the network edge) instead of central cloud servers (far from IoT devices) to improve overall system efficiency. However, offloading the whole application to a single MEC server (i.e. coarse-grained offloading) may cause an availability problem, which reduces the system’s overall performance. Fine-grained offloading is one possible solution, but it introduces further challenges, including subtask dependencies, per-task offloading decisions, and result integration. Having centralized control and a global view of the underlying elements is required to overcome these challenges, and Software Defined Networking (SDN) provides such capabilities. This thesis proposes an SDN-based dependency-aware partial offloading framework for IoT edge networks (SPOT). It uses Directed Acyclic Graph (DAG) based task partitioning to model complex IoT applications as dependent subtasks, enabling fine-grained offloading decisions. Leveraging SDN’s global view for both vertical offloading (device-to-edge) and horizontal offloading (edge-to-edge), SPOT facilitates dynamic flow scheduling that seamlessly integrates partial results while preventing network congestion. Extensive simulations demonstrate that SPOT significantly reduces task completion times, lowers the rate of disregarded tasks, and achieves higher resource utilization compared to conventional methods such as local execution or binary offloading—underscoring the potential of SDN-driven, dependency-aware offloading to meet the stringent latency requirements of emerging IoT applications.