Towards automatic performance optimization of networks using machine learning

A key principle for optimizing network performances is to understand the relationship between network topology, configuration parameters, and their influence on the behavior of network protocols. While an attractive approach to this problem is to use formal models of protocols in combination with mathematical optimization, such methods are often limited by either poor scalability or approximations of the models, leading to weak usability for realistic use-cases. In order to overcome those drawbacks, an emerging trend has been to use machine learning based techniques, which offer accurate predictions of the performance of network protocols. Our contribution in this paper is a general methodology which combines the statistical analysis of realistic data with an optimization algorithm in order to plan and optimize network topologies. One benefit of our approach is that it does not require prior knowledge on the studied topology and protocol. As an application example of our method, we address the challenge of placing virtual machines in a given topology with the general goal of optimizing the average bandwidth of TCP flows. In our numerical evaluation, our approach results in an overall average increase of the performance of TCP flows by more than 50%.