Couper: DNN Model Slicing for Visual Analytics Containers at the Edge

Applications incorporating DNN-based visual analytics are growing in demand. This class of data-intensive and latency-sensitive workloads has an opportunity to benefit from the emerging edge computing tier. However, to decouple the growing resource demand of DNN models, from the characteristics and resource limitations of the infrastructure elements available at the edge, new methods are needed to quickly slice the DNNs into appropriately-sized components, and to deploy those DNN slices to be executed on the edge infrastructure stacks. This paper presents Couper, a practical solution that provides for quick creation of slices of production DNNs for visual analytics, and enables their deployment in contemporary container-based edge software stacks. Couper is evaluated with 7 production DNNs, under varying edge configurations.