Joint Service Placement and Model Partitioning for Accelerating DNN Inference in Edge Intelligence Empowered Vehicle Networks

This paper investigates the joint service placement and model partitioning problem for accelerating DNN inference in edge intelligence empowered vehicle networks, considering the vehicle mobility and the caching capability of Road Side Unit (RSU). To avoid the degradation of DNN inference performance due to the cold start during the vehicle's service node switching, we propose three different service placement solutions, including processing locally, directly selecting one RSU as server node for collaborative offloading, and selecting one RSU as relay to maintain the connection with original server node. We model the RSU network as an undirected weighted graph and formulate an optimization problem to minimize the DNN inference latency by jointly optimizing the DNN service placement and DNN model partitioning. To solve this problem, we propose a Graph Neural Network Reinforcement learning-based service Placement and model Partitioning algorithm (GRPP). Initially, we extract the feature embedding of RSU nodes and obtain the DNN service placement solution based on Graph Neural Network (GNN). Furthermore, we design a Multilayer Perceptron (MLP) to generate the DNN model partitioning point. Specifically, the parameters of GNN and MLP are learned through an actor-critic based reinforcement learning structure. Finally, we generate the simulation model via traffic simulator Simulation for Urban MObility (SUMO) and give a comprehensive performance evaluation. Simulation results indicate that our proposed scheme can decrease the DNN inference latency by up to 64.51% compared with the benchmark.