Research On Partitioning And Distribution Of Convolutional Neural Network Models In Edge Environment

With the development of network communication technology and the arrival of the era of Internet of Everything,network edge devices and the amount of data are growing rapidly.The centralized processing model represented by cloud computing cannot meet the needs of massive edge data in terms of real-time performance,energy consumption,and privacy security.New applications such as smart driving and smart cities also promote the wider application of deep learning models in Internet of Things(Io T)scenarios.However,edge devices in the Io T environment have limited resources and cannot satisfy the resource requirements of deep learning models for computing,storage,and communications.Therefore,efficient model partitioning and distribution scheduling strategies are key to deploying and running deep learning models in resource-constrained edge clusters.In this paper,the deployment and operation scheme of the Convolutional Neural Network(CNN)model under the edge cluster of the Internet of Things are deeply studied.Based on the model partitioning and distributed collaborative inference method,the corresponding solution is proposed to reduce the memory footprint of the edge devices and the time of inference execution.The main research is summarized as follows:(1)For the resource-constrained edge environment of the Io T,the basic principles of model parallelism are analyzed,and an adaptive scheme based on the vertical fusion partition model is proposed.In the edge scenario,a single edge device usually cannot run and store the entire neural network model,and the existing division strategy cannot effectively adapt to the edge device.The proposed scheme can efficiently divide model according to the memory conditions of edge devices,and quickly adapt to edge devices by its self-adaptive ability.The experimental analysis shows that it can effectively reduce memory footprint,reduce intermediate feature data transmission overhead,and explore higher task parallelism.(2)For the partition task after adaptive partitioning,an efficient distributed collaborative inference strategy is proposed.For convolutional layers that require a lot of resources,based on the nature of convolution operations,there will be data overlaps between partition tasks after vertical division,resulting in a lot of redundant calculations.The data reuse scheme is adopted to analyze the dependence relationship caused by the reuse of data,and a method to quantify the dependence degree of partition tasks is proposed.Under the constraints of dependency and cost function,a distributed scheduling strategy for parallel execution of partitioned tasks is proposed.Through experimental analysis and evaluation,the proposed strategy can efficiently schedule partition tasks,reduce communication overhead and inference execution time effectively.