The DAG-Based Hybrid Deep Neural Network (DNN) Tasks Scheduling Algorithm (DHTS) in Real-Time Edge Systems

With the development of AI of things (AIoT), real-time edge systems play a key role in multiple fields such as autonomous driving, real-time road monitoring, avionics, and are capable of performing a large number of hybrid deep learning (DL) inference tasks. At present, more and more embedded real-time edge systems use heterogeneous CPU and GPU cores to improve the performance of system processing tasks. However, the research on hybrid deep neural network (DNN) inference tasks with DAG dependency is not comprehensive, and the factors such as system resource utilization and real-time response of task scheduling are not fully considered. In this paper, the DAG-based Hybrid DNN Tasks Scheduling algorithm (DHTS) is proposed. In this algorithm, the joint optimization goal of minimizing the resource occupancy and energy utilization of the real-time edge system is established, and a hybrid DNN task scheduling scheme with DAG dependence is designed. This paper evaluates the performance of DHTS algorithm through simulation experiments. The experimental results show that DHTS algorithm can realize efficient DAG-dependent hybrid DNN task scheduling, make full use of heterogeneous computing resources, and minimize resource occupancy and energy consumption of real-time edge system.