Predictive GPU-based ADAS management in energy-conscious smart cities

Abstract

The demand of novel IoT and smart city applications is increasing significantly and it is expected that by 2020 the number of connected devices will reach 20.41 billion. Many of these applications and services manage real-time data analytics with high volumes of data, thus requiring an efficient computing infrastructure. Edge computing helps to enable this scenario improving service latency and reducing network saturation. This computing paradigm consists on the deployment of numerous smaller data centers located near the data sources. The energy efficiency is a key challenge to implement this scenario, and the management of federated edge data centers would benefit from the use of microgrid energy sources parameterized by user's demands. In this research we propose an ANN predictive power model for GPU-based federated edge data centers based on data traffic demanded by the application. We validate our approach, using real traffic for a state-of-the-art driving assistance application, obtaining 1 hour ahead power predictions with a normalized root-mean-square deviation below 7.4% when compared with real measurements. Our research would help to optimize both resource management and sizing of edge federations.