Research On DNN Task Partition And Offloading In Vehicular Edge Network

In recent years,with the development of technologies such as artificial intelligence and sensors,vehicles provide diverse applications and services,such as obstacle recognition and lane detection.Most of these applications are latency-sensitive,and are generally based on deep neural networks(DNNs),which are characterized by high accuracy and high computational and energy resource consumption.Due to the limited computing resources of vehicles,the low latency requirements of these applications cannot be met.Vehicular Edge Computing(VEC)can enhance the computing power of vehicles by sinking computing resources to Road Side Units(RSUs)on both sides of the road and make up for the shortage of computing and energy resources.However,due to the mobility of vehicles and the uneven load of edge servers,as well as the diversity of DNN topological structures(such as chain structure and DAG structure,etc.),how to offload DNN tasks to edge servers to optimize latency and reduce consumption of vehicle resources is still a challenging problem.Therefore,this thesis focuses on how to partition and offload tasks for different DNN topology models in the VEC environment to improve the performance of vehicle DNN task processing.The main work and innovations of this thesis are as follows:(1)First,this thesis takes the VGG11 model as an example to study and analyze the calculation time,network transmission time,end-to-end delay,and energy consumption of the DNN task on the vehicle and edge server processing through experiments and illustrating feasibility of DNN task partition and offloading in the VEC environment.(2)This thesis proposes a DPAO method for the chained DNN model,which balances the task computation time and data transmission time through vehicle-edge collaboration and edgeedge collaboration,and dynamically offloads tasks according to the difference in edge server load.The experimental results show that the average task completion time is reduced by 1.9 times compared with directly offloading to the edge server;compared to only computing locally in the vehicle,the vehicle energy consumption is reduced by 94%.(3)Aiming at the DAG-style DNN model,this thesis proposes the DCAO method.The partition problem of the DNN model is transformed into a maximum flow minimum cut problem.Based on the existing research,the network flow graph is further expanded to reduce the time complexity of the partition algorithm.For the case that the network flow graph has negative capacity,this thesis converts all of them into positive numbers,and then solves it through the traditional minimum cut algorithm.The experimental results show that the average task completion time is reduced by 47% compared with direct offloading to the edge server;compared to only computing locally in the vehicle,the vehicle energy consumption is reduced by 56%.