Real-time DNN Model Partitioning for QoE Enhancement in Mobile Vision Applications

As deep learning technology advances, mobile vision applications such as augmented reality or autonomous vehicles are widespread. The quality of experience (QoE) of such applications highly depends on hardware specification of mobile device, dynamic service requests, stochastic network status and characteristics of DNN model. In this paper, we propose an algorithm called RT-DMP that jointly optimizes DNN model partitioning and process/network resources adapting to system dynamics by leveraging virtual queue-based Lyapunov optimization framework. The RT-DMP jointly makes decisions on (i) partition point between a mobile device and an MEC server, (ii) mobile GPU clock frequency, and (iii) transmission rate through the wireless network every time slot. We theoretically show that RT-DMP optimally strikes the balance among three QoE metrics that are energy consumption, throughput and end-to-end latency, which has not been addressed in existing studies. Finally, we demonstrate the performance and feasibility of RT-DMP via trace-driven simulations and real testbed based on Nvidia Jetson TX2 and a high-end MEC server.