Predictive analytics for fog computing using machine learning and GENI

摘要

Fog computing is a rapidly emerging paradigm concerned with providing energy- and latency-aware solutions to users by moving computing and storage capabilities closer to end users via fog networks. A major challenge associated with such a goal is ensuring that forecasts about network quality are not only accurate but also have small operational overhead. Machine Learning is a popular approach that has been used to model network parameters of interest. However, due to the small amount of public datasets and testbeds available, designing reproducible models becomes cumbersome and more likely to under-perform during deployment. For these reasons, we seek to design an exploratory testbed for benchmarking the forecasting strength of a suite of supervised learning models aimed at inferring network quality estimates. To create a realistic fog computing sandbox, we deployed an image processing ensemble of services in the GENI infrastructure. The nodes in GENI have varying hardware specifications for the purpose of generating compute-intensive workloads on heterogeneous systems. Our experimental results suggest that machine learning can be used to accurately model important network quality parameters and outperforms traditional techniques. Moreover, our results indicate that the training and prediction times for each model is suitable for deployment in latency-sensitive environments.