Dynamic Task Division and Allocation in Mobile Edge Computing Systems: A Latency Oriented Approach via Deep Q-Learning Network

With the rapid development of Internet of Things (IoTs), various sensors are deployed to collect different physical information. Smart surveillance is one of applications by analyzing the real-time video generated by camera sensors. However, due to the limited computing capability of camera sensors, running video analysis models (e.g., AlexNet and YOLO3) on camera sensors directly consumes a lot of computing time. In addition, transferring video to the remote cloud suffers a long-distance transmission latency. Fortunately, edge computing has been considered as a promising solution for enabling computation-intensive yet latency-sensitive applications at resource-constrained devices. Thanks to edge computing, camera sensors can upload video to different edge servers employed at the edge of networks for processing. Moreover, the lightweight Kubernetes for edge computing, i.e., K3s, enable a fine-grained task division and parallel computing. In this paper, we consider a heterogeneous edge cooperative video analysis, i.e., face recognition, with the objective of minimizing the processing latency. Specifically, we use a Deep Q-Learning network (DQN) to dynamically adjust the size of pieces video allocated to different edge servers connected via wireless networks. In addition, to improve the resource utilization of edge servers and reduce the processing latency, each edge server further divides the received video into multiple segments that are processed by different containers in parallel. To validate the effectiveness of our scheme, we implement a small-scale prototype system and conduct numerous experiments. Experimental results show that our proposed algorithm outperforms the other four schedule schemes by testing on the tasks of face recognition and pose recognition.