Parecon: Enhancing Mobile AI Vision Apps through DNN Partition and Resolution Control

초록

Vision applications using deep neural network (DNN) models are increasingly prevalent in mobile devices like autonomous vehicles, drones, and smartphones. The quality of experience (QoE) for these applications is affected by the hardware performance of the mobile devices, fluctuating network conditions, and the characteristics of the DNN model. In this paper, we introduce the Parecon algorithm, which optimizes DNN model partitioning and frame resolution based on stochastic optimization adapting to both internal and external system dynamics. Parecon dynamically determines (i) DNN model partition point between the mobile device and the MEC server, (ii) resolution of the input frame, and (iii) number of processed frames per second, which have not been jointly addressed in previous studies. We theoretically demonstrate that Parecon optimizes three key QoE metrics, i.e., end-to-end (E2E) latency, accuracy, and throughput. Furthermore, we validate the effectiveness and superiority of Parecon over existing algorithms through trace-driven simulations and testbed implementation based on an embedded device and a high-end GPU server. © 2025 IEEE.

키워드