DNN Inference Acceleration Based on Adaptive Task Partitioning and Offloading in Embedded VEC
Chunlin Li, Sen Liu, Kun Jiang, Mengjie Yang, Zihao Zhang, Bingxin Wang, Liang Zhao, Chen Chen, Shaohua Wan
Abstract
As a distributed embedded system, vehicular edge computing (VEC) completes various complex Deep neural network (DNN) tasks through network collaboration and communication. However,due to the limited computing power of vehicle processors, vehicles cannot handle increasingly complex DNN tasks. To accurately estimate the execution latency of each layer across different DNN models on heterogeneous devices, we proposed the Extreme Gradient Boosting Tree (XGBoost) algorithm to predict DNN task inference latency. Furthermore, we proposed partitioning and offloading algorithms for both chained DNN tasks and Directed Acyclic Graph (DAG)-type DNN tasks, addressing their unique computational characteristics. For chained DNN tasks, we employ a linear search to determine optimal partitioning points based on predictions from the DNN latency prediction model. For the partitioning and offloading of DAG-type DNN tasks, we construct it as a minimum cut problem under the network flow graph and propose a DNN task partitioning and offloading algorithm based on the highest label pre-stream push (HLPP) algorithm to effectively reduce the cost of task partitioning and offloading. Finally, we used an experimental vehicle equipped with Raspberry and a RSU equipped with Jetson Nano to verify the results. The experiment shows that the DNN latency prediction model based on the XGBoost we proposed can effectively improve the latency prediction accuracy of DNN layer-by-layer execution. At the same time, the division and offloading algorithms for different types of DNN inference tasks can achieve higher task completion rate, lower latency, and lower energy consumption.