Delay-Energy Tradeoff for Intelligent Online Partial Offloading in Mobile Edge Computing

Xianlong Jiao, Yicheng Zhao, Yilang Feng, Songtao Guo, Xianzhang Chen, Mingyan Li, Wei Lou

Abstract: Task offloading is a pivotal technology for mobile edge computing (MEC). However, limited channel resources make it challenging to support all user devices (UDs) offloading tasks to edge servers. Therefore, making rational task offloading decisions is paramount to enhancing the MEC system performance. Specifically, this study focuses on the online partial offloading problem to minimize the delay-energy trade-off. The presence of inter subtask dependencies, time-varying wireless channels, dynamically varying residual energy, and randomly changing task data makes this problem very challenging, and traditional algorithms are hard to solve this problem. To effectively address this challenge, we combine the scheduling based on the directed acyclic graph (DAG) with the deep reinforcement learning (DRL) technology, and introduce an intelligent online partial offloading algorithm named IOPO. IOPO utilizes the DAG to obtain the parallelable subtask set, learns from experience through an elaborate DRL framework and generates offloading decision matrices. The DRL framework ensures feasible actions in real time. Numerical results show that, IOPO surpasses benchmark algorithms in terms of delay-energy trade-off and works well in dynamic MEC settings with fast convergence and low running time.