Provisioning Deep Learning Inference on a Fog Computing Architecture

As a consequence of the exponential growth of current technology and, by nature, information digitization, there is an increasing number of final devices on the Internet of Things (IoT). To address these requirements, it is appropriate to access new capabilities of information technologies (IT) and operational technologies (OT) and have the potentialities offered by the Cloud, but from an environment close to the production plant, which implies a reduction in broadband costs and low dependence on latency. With this approach, this study presents a Fog Computing (FC) alternative implemented for object recognition through Artificial Intelligence (AI). The architecture proposed consists of deep neural networks (DNNs) processing nodes, the main Fog node for Docker-based provisioning and Kubernetes orchestration; in addition, Prometheus was used for collecting dynamic metrics, such as central processing unit (CPU), random access memory (RAM) and power absorbed, and afterward, Grafana was used for the analysis and visualization of their trends. After experimentation, the Fog architecture was obtained by balancing the workload in the two processing nodes, achieving an improvement of more than 50% in the provisioning and orchestration processing time compared to an architecture constituted by a single node.