Characterizing the Execution of Deep Neural Networks on Collaborative Robots and Edge Devices
Matthew L. Merck, Bingyao Wang, Lixing Liu, Chunjun Jia, Arthur Siqueira, Qiusen Huang, Abhijeet Saraha, Dongsuk Lim, Jiashen Cao, Ramyad Hadidi, Hyesoon Kim · Proceedings of the Practice and Experience in Advanced Research Computing on Rise of the Machines (learning) · 2019
Edge devices and robots have access to an abundance of raw data that needs to be processed on the edge. Deep neural networks (DNNs) can help these devices understand and learn from this complex data; however, executing DNNs while achieving high performance is a challenge for edge devices. This is because of the high computational demands of DNN execution in real-time. This paper describes and implements a method to enable edge devices to execute DNNs collaboratively. This is possible and useful because in many environments, several on-edge devices are already integrated in their surroundings, but are usually idle and can provide additional computing power to a distributed system. We implement this method on two iRobots, each of which has been equipped with a Raspberry Pi 3. Then, we characterize the execution performance, communication latency, energy consumption, and thermal behavior of our system while it is executing AlexNet.