Collaborative Mobile Edge Computing Through UPF Selection

The distributed deployment and the relatively limited resource of one edge node make it quite challenging to effectively manage resources at the edge. Inappropriate scheduling may result in a quality of service deterioration and brings significant cost. In this paper, we propose a per-user level management mechanism for joint scheduling of user requests and container resources at the edge and study how to minimize average cost as well as satisfy delay constraints. The cost model of the system consists of operating cost, switching cost and delay violation cost. The key idea is to deploy a deep reinforcement learning-based scheduler in the core network to conduct joint network and computation management. To evaluate the performance, we build a test bed namely MiniEdgeCore that contains a full user plane protocol stack and deploy a real-time video inference application on it. A real-world dataset is used as the workload sequence to conduct experiments. The results show that the proposed method can reduce average costs effectively.