Abstract:Modeling 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 such complex systems. One drawback of this method is that high-level features are generally used – which require full understanding of the network protocols to be chosen, correctly engineered, and measured – and the approaches are generally limited to a given network topology. In this article, we present DeepComNet, an approach to address the challenges of working with machine learning by using lower-level features, namely only a description of the network architecture. Our main contribution is a method 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. To evaluate the potential of DeepComNet, we apply our approach to the tasks of predicting the throughput of TCP flows and the end-to-end latencies of UDP flows. In both cases, the same base model is used. Numerical results show that our approach is able to learn and predict performance properties of TCP and UDP flows with a median absolute relative error smaller than 1%, outperforming related methods from the literature by one order of magnitude.

DeepComNet: Performance evaluation of network topologies using graph-based deep learning