Edge Intelligence?Research On Cloud-Edge-End DNN Collaborative Inference Acceleration Technology

The Internet of Everything has brought about the explosive growth of edge applications and data close to mobile terminals,and has also enabled artificial intelligence to have richer and more humanized application scenarios.Pushing artificial intelligence to the edge,using the data and computing resources of the edge to fully release its potential,has become the most popular solution at present,and the concept of edge intelligence has been born from this.Although deep learning improves the accuracy of various intelligent applications,it also brings higher computing requirements.Placing DNN inference tasks in the cloud will result in larger network data transmission,but pushing the task completely to the device or the edge also means heavier computing resources are occupied.Therefore,this subject mainly studies cloud edge-end DNN collaborative reasoning acceleration technology for edge intelligence.First of all,this article analyzes the existing model partitioning algorithms,based on the fine-grained changes in the output data and calculations of each layer of DNN,and discusses the performance and function problems of the more complex DAGstyle DNN model partitioning.On the one hand,the design drawing compression method solves the performance problem of DNN division,and the performance of the algorithm is verified by experiments based on multiple sets of common DNN models;on the other hand,it solves the problem of division defects existing in partial parallel units.Finally,on the basis of the two-layer partition algorithm,a three-layer twostage model partition scheme for the cloud side end is further designed.After that,a cloud-side-end collaborative reasoning framework was designed and implemented,and multi-user analysis and solving the problems of reasoning task congestion,resource preemption and training task scheduling were solved.On the basis of the model partition algorithm,a layered training method is further designed to improve the efficiency of model update and the accuracy of reasoning.Finally,based on the real edge data set,a large number of comparative experiments are designed to verify the efficiency of the algorithm and the system.The experimental results show that compared with the edge and the central cloud executing the entire DNN model,the reasoning delay of the three-layer division scheme in this paper is up to 4.99 times and 2.1 The throughput is increased by 8.7times and 3.4 times.Compared with the two-layer division algorithm DADS,the reasoning delay is increased by up to 15%,and the division time is reduced to only1%.