移动边缘网络中深度学习任务卸载方案

DEEP learning task offloading scheme in mobile edge network

为了缓解网络带宽的压力、降低网络延迟,克服移动设备资源匮乏等问题,推动深度学习应用在移动终端的部署,提出一个基于移动边缘计算的深度学习任务卸载方案。基于深度神经网络专用加速芯片Eyeriss的架构,对深度学习任务的计算功耗进行建模,提出了一个基于混合1/2范数诱导的三阶段组稀疏波束成形(group sparse beamforming,GSBF)框架,通过对计算任务优先级的精心设计,尽可能地删除基站端冗余的计算任务,实现对整体网络功耗(包括发送功率损耗和计算功率损耗)的优化。针对该框架,提出了一个加速优化方案。仿真实验表明,在该场景下,所提出的框架在优化整体网络功耗方面具有显著优势,而加速算法可以进一步提升框架的性能。

To alleviate the pressure of network bandwidth,reduce the network delay,overcome the shortage of mobile equipment resources,and promote the deployment of deep learning application on mobile terminals,we propose a deep learning task offloading scheme based on mobile edge computing.Firstly,based on the architecture of Eyeriss,an acceleration chip specially designed for deep neural network,the computational power of deep learning task is modeled.Then,a three-stage group sparse beamforming(GSBF)framework based on hybrid 1/2 norm inducing is proposed.By carefully designing the computing task priority,the redundant computing tasks at the base station end are deleted as much as possible,and the optimization of the whole network power consumption including transmission power consumption and calculation power consumption is realized.Besides,an accelerated optimization scheme is proposed for this framework.Simulation experiments show that the proposed framework has significant advantages in optimizing the entire network power consumption under this scenario,and the accelerated scheme can further improve the performance of the framework.