Performance Evaluation of Network Topologies using Graph-Based Deep Learning

Understanding the performance of network protocols and communication networks generally relies on expert knowledge and understanding of the different elements of a network, their configuration and the overall architecture and topology. Machine learning is often proposed as a tool to help modeling complex protocols. One drawback of this method is that high-level features are generally used -- which require expert knowledge on the network protocols to be chosen, correctly engineered, and measured -- and the approaches are generally limited to a given network topology.

In this paper, we propose a methodology to address the challenge of working with machine learning by using lower-level features, namely only a description of the network architecture. Our main contribution is an approach for applying deep learning on network topologies via the use of Graph Gated Neural Networks, a specialized recurrent neural network for graphs. Our approach enables us to make performance predictions based only on a graph-based representation of network topologies. We apply our approach to the task of predicting the throughput of TCP flows. We evaluate three different traffic models: large file transfers, small file transfers, and a combination of small and large file transfers. Numerical results show that our approach is able to learn the throughput performance of TCP flows with good accuracies larger than 90%, even on larger topologies.