On Estimating Communication Delays using Graph Convolutional Networks with Semi-Supervised Learning

In large-scale communication networks consisting of many end hosts and routers, accurate acquisition, measurement, and estimation of communication delays between node pairs are essential for providing high-quality communication services. In several QoS-related performance metrics such as metrics for efficiency, metrics for availability, metrics for reliability, communication delay is one of the key metrics to realize several traffic control mechanisms. Conventional instrumentation and measurement techniques are suitable when the size of the network to be measured is relatively small, or when the number of node pairs to be measured is relatively small. However, in evolving and complex networks, it is not trivial to acquire, measure, and estimate the communication quality at a huge number of routers and end hosts. In this paper, as an initial step toward the realization of estimating communication quality (especially communication delays between node pairs) in a large-scale network, we investigate the potential of graph neural networks with semi- supervised learning for estimating communication delays between node pairs. Our findings include that the average relative error of estimated communication delays is around 10-35% depending on the fraction of the number of measurement nodes, and that communication delays for large-scale networks can be estimated with a high accuracy even if the fraction of the number of measurement nodes is not so large.