1（State Key Laboratory for Strength and Vibration of Mechanical Structures，Xi'an Jiaotong University，Xi’an 710049，China）2（Beijing Aeronautics Techno Research Center，Beijing 100076,China）3（School of Computer Science，Northwestern Polytechnical University，Xi’an 710129，China）
Abstract：In recent years，neural networks have made great achievements in image classification and object detection.Many scholars began to study how to deploy neural networks efficiently on heterogeneous edge devices and proposed the idea of dividing data into Hot/Cold-Class.However，due to the great difference in processing performance between front and back ends(CPU-GPU)of heterogeneous edge devices，the traditional training method will lead to the resource mismatch.This paper presents a task partition scheduling algorithm basedon load balancing.By accurately estimating the real-time load of the front-end CPU of heterogeneous edge devices overaperiodtimein the future，this method can allocate resources dynamically，and effectively solve the problem of unbalanced utilization of resources at the front and back ends of heterogeneous edge devices.The experiment shows that this method can effectively divide resources,and in the case of satisfying load balancing，energy consumption is reduced by 31.4% compared with traditional methods.
2021, Vol. 42 
