An intelligent scheduling framework for DNN task acceleration in heterogeneous edge networks

Abstract

With the upgrade of hardware architecture and device capacities, many accelerator-based hardware platforms have been widely deployed in Mobile Edge Computing (MEC) environments. The execution time of many computation-intensive applications (e.g., face recognition and pedestrian detection) can be significantly reduced when deployed on these heterogeneous devices. Moreover, thanks to the popularity of Deep Learning (DL), most terminal applications are integrated with Deep Neural Networks (DNN) and can be divided into interdependent tasks. The structure of these applications can be represented as the Directed Acyclic Graph (DAG). Therefore, it is critical to seek the optimal scheduling order and execution placement of tasks according to the acceleration effects of edge servers and the task dependency. However, conventional scheduling strategies focus on the short-term performance, potentially leading to service quality degradation in the long term. Besides, many studies use Deep Reinforcement Learning (DRL) algorithms to seek a long-term optimal scheduling strategy but ignore the device acceleration and the task dependency. Furthermore, training a well-performed DRL agent is time-consuming, and the large scale of trial-and-error will take up tremendous computation and storage resources. In this paper, we model the scheduling process as a Markov Decision Process (MDP) and design an adaptive scheduling framework for task acceleration. Fully considering the data dependencies, resource conditions, and network conditions, the proposed scheduling algorithm called Meta-AC uses policy gradient combined with meta-learning to minimize the average task delay and the ratio of time-out tasks. As a hierarchical DRL algorithm, Meta-AC uses meta data to learn directed exploration strategies in the high-level agent, improving the learning efficiency from experience samples. Extensive simulations demonstrate the superiority of the proposed method over the counterpart methods.