移动边缘计算环境中基于能耗优化的深度神经网络计算任务卸载策略

Energy efficient computing task offloading strategy for deep neural networks in mobile edge computing

深度神经网络因其强大的数据分析功能而被广泛使用在移动智能应用中,然而其计算任务的复杂性给计算能力与电池容量均有限的终端设备带来了巨大的挑战。若将深度神经网络中的计算任务完全卸载到云端,则会产生较高数据传输时延。移动边缘计算因其低时延、分布式、位置感知的优势能有效解决深度神经网络的时延和能耗问题。为了在保证用户时间约束的同时,充分优化终端设备能耗,建立了移动边缘计算环境中深度神经网络计算任务卸载的时间和能耗评价模型,基于该模型提出了移动边缘计算环境中基于能耗优化的深度神经网络计算任务卸载策略。该策略以神经网络层为单位,将深度神经网络中的计算任务进行拆分,在计算任务卸载时,对移动边缘计算环境中多重计算资源进行综合考虑。最后,提出了移动边缘环境中基于多重资源任务卸载的粒子群调度算法,该算法能在满足时间约束的同时,充分优化终端设备能耗。实验表明,与已有的3种任务卸载策略相比,新策略对应的粒子群调度算法所得适应度值最优,满足时间约束下,终端设备的能耗值最低。

Deep Neural Networks(DNNs)are widely used in mobile smart applications given their powerful data analysis capability.However,the complexity of their computing tasks brings a big challenge to end devices which are limited by their computing power and battery capacity.If the computing tasks in DNNs are offloaded completely to the cloud,the data transmission latency can be very significant.In contrast,with the advantage of low latency,distributed and location awareness,mobile edge computing can solve latency and energy-constrained problems in DNNs effectively.To optimize the energy-consumption of end devices with a user deadline constraint,the model of time and energy-consumption of computing tasks offloading in DNNs was established in the mobile edge computing.An energy efficient task offloading strategy for deep neural networks computing task in mobile edge computing was proposed.This strategy taken search layer of DNNs as a basic unit to divide computing task in DNNs and consider multiple computation resources synthetically in mobile edge computing environment during task offloading.A particle swarm optimization based task scheduling algorithm with multiple-resource task offloading was proposed,which could effectively optimize the energy-consumption of end devices under response time constraints.The experimental results showed that the proposed strategy could achieve the best fitness value compared with three other existing offloading strategies,and the end devices had the lowest energy consumption under response time constraints.