Jisuanji kexue (Nov 2022)

Workload Characteristics Based Performance Optimization for Edge Intelligence

  • HU Zhao-xia, HU Hai-zhou, JIANG Cong-fengand WAN Jian

DOI
https://doi.org/10.11896/jsjkx.211000067
Journal volume & issue
Vol. 49, no. 11
pp. 266 – 276

Abstract

Read online

Edge intelligence refers to a form of service that uses artificial intelligence algorithms to provide data analysis capabilities for network edge devices.However,the edge computing environment is more complex and changeable than cloud computing.There are many problems in the process of building edge intelligence,such as the lack of quantitative evaluation standards,heterogeneous computing platforms,complex network topologies,and changing user needs.Among them,the more prominent is the contradiction between the high resource demand of the algorithm model and the low resource reserve of edge devices.Machine lear-ning is the main workload of edge intelligence.It requires a lot of computing resources.However,the computing resources of edge devices are limited,and the supply and demand between the two do not match.The deployment and optimization of edge intelligent load has become a problem.Therefore,in response to the problem of edge intelligent load performance optimization,this paper proposes cloud-edge collaborative inference(CECI) based on load characteristics,which is optimized for different machine learning loads in terms of model selection,batch adaptive adjustment and cloud-side collaboration.In terms of model selection,a model adaptive selection strategy based on target weights is used to comprehensively weigh the effects of multiple performance optimization targets under multiple constraints.In the aspect of batch adaptive adjustment,a batch adaptive adjustment algorithm based on overhead feedback is proposed,so that the model can achieve better performance at runtime.In terms of cloud-side collaboration,a cloud-side collaboration strategy is designed by combining network status and user delay requirements to achieve the effect of dynamic utilization of cloud computing resources.Experimental results show that compared with cloud intelligence,the edge intelligence based on load characteristics proposed in this paper can reduce program running time by 50.79%,reduce system energy consumption by 42.46%,and improve model accuracy by 4.52%.

Keywords