Pao-Ding: Accelerating Cross-Edge Video Analytics via Automated CNN Model Partitioning

Cross-edge video analytics is a technology that involves collaborative processing of video data among multiple edge devices. To handle increasingly complex video analytics tasks on edge devices with higher frame rates or improved accuracy, cross-edge video analytics leverages model partitioning techniques to divide Convolutional Neural Network (CNN) models into multiple sub-models deployed across different computing nodes. This approach enables the efficient utilization of heterogeneous edge computing resources for compute-intensive CNN inference. Nonetheless, several challenges persist in cross-edge video analytics. These include the partitionability of CNN models requiring a lot of manual work, excessive transmission of intermediate features leading to performance degradation, and high computational complexity in partitioning decision-making. To tackle these challenges, this paper proposes Pao-Ding, a framework to accelerate cross-edge video analytics. Pao-Ding employs gradient information to automatically parse the structure of CNN models, thereby enabling partitionability for chain, simple Directed Acyclic Graph (DAG), and complex DAG models. By incorporating a skip-layer design, Pao-Ding is able to compress redundant intermediate features more accurately. Additionally, a layer-pruning-based algorithm is proposed in Pao-Ding which effectively reduces the computational overhead associated with partitioning decision-making. Experimental results on real hardware show that Pao-Ding supports automated model partitionability for 67 CNN models with a model coverage rate of 94.37%. Compared to state-of-the-art, Pao-Ding reduces total prediction error by 47.43%, lowers the partitioning decision time by 37.5%–58.35%, and decreases the average video frame processing time by 5.75%–22.18%. Furthermore, a cross-edge video analytics simulation platform developed using Docker and Pumba verifies the robustness of Pao-Ding under various conditions involving different number of nodes, computing capacities, and network environments.