Performance Analysis of a CNN Counting Application for Fog and Cloud Computing

Abstract

Data extraction from surveillance videos is an important subject, not only because of the amount of data generated, but also because it is hardly ever processed. Advances in Edge and Fog computing could allow having a processing closer to source of the video. However, streaming video flows to the Cloud seems feasible too. In the context of an automatic counting application, using Convolutional Neural Networks (SSDMobilenet, GoogleNet) for detection and classification, this work address the following question: How many flows can a server handle without downgrading acceptable performance? This article presents the analysis of performance of the counting application running on the Cloud and on the Fog. Analysis include consumption of: network, RAM, CPU, and GPU. These tests allow a better sizing of the hardware requirements to deploy the counting application. Different tests are defined to isolate specific case behavior for regular video’s resolution (1920x1080@20–30fps). Results indicate that a restricted number of simultaneous flows is possible, even when GPU is used; i.e. 5–7 flows. Performance is even worse for a CPU only scenario, suggesting additional processing techniques should be used to reduce load.