A Machine Learning approach for dynamic selection of available bandwidth measurement tools

Available bandwidth is a vital parameter for understanding network status. A huge number of tools have been proposed in literature, including general ones as well as tools specialized for specific network scenarios. In this plethora of possibilities, an expert user is currently required to select the best one according to the specific operating settings, in order to achieve accurate results without disturbing the existing traffic. In this paper we propose the use of automatic decision systems based on machine learning to substitute the expert user and choose the right tool for the current scenario. To verify if this is a viable solution, we created a custom dataset and tested different decision systems including a simple, threshold-based one and four algorithms based on machine learning: k-nearest Neighbors, Random Forest, Support Vector Machine, and Long Short-Term Memory networks. We used different features including CPU, memory, and bandwidth and verified that the decision systems based on machine learning achieve very good performance and can be considered as a promising solution for this important problem.