Adaptive Task Distribution in Edge Networks using SDN and DRL

Abstract

A rise in computation-intensive mobile applications, such as augmented reality, real-time video analytics, and autonomous systems, has increased demand for low-latency, computationally intensive solutions. Traditional centralised cloud-based solutions are powerful but unable to meet low-latency requirements due to the distance involved. Therefore, there is a shift towards Mobile Edge Computing (MEC), which reduces execution latency and energy consumption by offloading tasks to proximate servers. Nevertheless, most current offloading studies either rely on binary task models or overlook the internal dependency architecture of programs. Thus, it limits their scheduling effectiveness. Furthermore, the combination of partial offloading, task dependency awareness, and Software-Defined Networking (SDN) integration remains largely unexplored. This thesis proposes a Double Deep Q-Network (DDQN)-based partial offloading framework for dependency-aware task scheduling in SDN-enabled MEC networks. The proposed agent operates on a dual-graph state representation comprising the task dependency graph and the network resource graph. It encourages per-subtask placement decisions that jointly account for computational structure and resource availability. The framework is evaluated against DQN, A2C, PPO, Random, and Binary baselines across several dimensions: workload size, subtask density, and network topology in normal and scarce resource conditions. Results show that the DDQN agent reduces total execution time by up to 58\% compared to uninformed baselines, while maintaining near-zero task discard rates under scarce resource conditions where binary offloading degrades sharply.