Collaborative cloud-edge-end task offloading with task dependency based on deep reinforcement learning

With the explosive growth of the Internet of Things (IoT), IoT devices generate massive amounts of data and demand, which poses a huge challenge to IoT devices with limited CPU computing capability and battery capacity. Due to the dependency relationship in complex applications and network environments, effective offloading in such scenarios is complicated. In this paper, we address the problem of computation offloading with task dependency in the cloud–edge-end collaboration scenario including multi-user, multi-core edge servers, and a cloud server. We model multiple task dependencies using the directed acyclic graph (DAG) and formalize the offloading problem as a multi-objective mixed integer optimization problem. To solve this problem, a Task Priority and Deep Reinforcement learning-based Task Offloading algorithm (TPDRTO) is proposed. The task offloading decision is represented as a Markov Decision Process (MDP). Meanwhile, based on the task priority, an optimized Deep Reinforcement Learning (DRL) method with action-mask is proposed to leverage the computing resources of the cloud and edge servers and obtain the optimal policies of computation offloading. Experimental results show that the TPDRTO algorithm can effectively tradeoff and reduce the average energy consumption and time delay of IoT devices.