基于差分进化的推断任务卸载策略  

Inference Task Offloading Strategy Based on Differential Evolution

作　　者：王瑄 毛莺池[1] 谢在鹏 黄倩[1] WANG Xuan;MAO Ying-chi;XIE Zai-peng;HUANG Qian(School of Computer and Information,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学计算机与信息学院,南京211100

出　　处：《计算机科学》2020年第10期256-262,共7页Computer Science

基　　金：国家重点研发课题(2018YFC0407105);国家自然科学基金重点项目(61832005);中央高校科研业务费(2017B20914);华能集团重点研发课题(HNKJ17-21)。

摘　　要：卷积神经网络(Convolutional Neural Network,CNN)作为深度学习的重要技术,已被广泛应用在移动智能应用中。针对CNN推断任务高内存、高计算量的需求,现有解决方案多将任务卸载到云上执行,难以适应时延敏感的移动应用程序。为解决上述问题,提出了一种基于改进差分进化算法的CNN推断任务卸载策略,它采用端云协作模式将计算任务部署在云和边缘设备之间。该策略研究了成本约束下最小化时延的任务卸载方案,将CNN推断过程转化为任务图并将其构建为0-1整数规划问题,利用改进二进制差分进化算法高效求解最佳卸载决策。实验结果表明,在给定费用约束下,与移动端推断和云推断方案相比,所提策略将任务响应时间平均缩短了33.60%和6.06%。As an important technology of deep learning,convolutional Neural Network(CNN)has been widely used in intelligence applications.Due to the demand of CNN inference task for high computer memories and computation,most of the existing solutions are to offload tasks to the cloud for execution,which are hard to adapt to the time-delay sensitive mobile applications.To solve the above problem,this paper proposes a CNN inference task offloading strategy based on improved differential evolution algorithm,which can efficiently deploy computing tasks between cloud and edge devices using end-cloud collaboration mode.This strategy studies the task unloading scheme that minimizes the time delay under cost constraint.transforms the CNN inference process into a task graph and constructs it into a 0-1 integer programming problem,and finally uses the improved binary differential evolution algorithm to solve the problem so as to infer the optimal offloading policy.The experimental results show that,compared with mobile inference and cloud inference schemes,averagely,the proposed strategy can reduce the task response time by 33.60%and 6.06%respectively with cost constraints.

关 键 词：卷积神经网络 移动云计算 计算卸载 协同推断 差分进化算法 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]