Joint DNN Partitioning and Multi-User Task Offloading in Multi-Server MEC Systems via Lyapunov Optimization

The rapid proliferation of multimedia applications, propelled by 5G technology, necessitates efficient data processing in resource-constrained terminal devices. Mobile Edge Computing (MEC) systems present a viable solution by offloading computational tasks to edge servers. However, existing methodologies typically neglect the structural intricacies of deep neural network (DNN) tasks and fail to accommodate multi-server environments. This paper introduces a DNN partitioning and task offloading strategy within a dynamic multi-user, multi-server MEC framework, with the objective of minimizing power consumption while adhering to stringent delay constraints. By employing Lyapunov optimization and formulating the problem as a multi-knapsack optimization task, the proposed algorithm enables terminals to locally process the initial layers of DNNs while offloading subsequent layers to edge servers. Simulation results demonstrate that the proposed approach significantly reduces power consumption and more effectively satisfies delay constraints compared to traditional methods.