Path-Link Graph Neural Network for IP Network Performance Prediction

Dynamic resource provisioning and quality assurance for the plethora of end-to-end slices running over 5G and B5G networks require advanced modeling capabilities. Graph Neural Networks (GNN) have already proven their efficiency for network performance prediction. GNN architecture matches well the structures usually met in communications networks. In this paper, the focus is on the IP transport network as one of the end-to-end 5G architecture domains. The recently published RouteNet GNN is taken as a reference and starting point for our study. RouteNet performance is verified by a new implementation in the PyTorch ML library. Next, an alternative Path-Link neural network (PLNet) architecture is proposed and evaluated. After hyper-parameter tuning for both models, the results show that PLNet and RouteNet achieve a similar accuracy level. The advantage of PLNet is in parallel architecture. It is demonstrated that its inference speed is not sensitive to the length of the network's paths.