Prediction Augmented Segment Routing

Abstract

With the increasing success of machine learning based approaches for prediction problems, there has been recent effort in improving the performance of online algorithms by augmenting them with machine learning predictions. Since machine learning predictions typically do not offer any performance guarantees, the new approach has to take into consideration the possibility that the machine learning prediction can be inaccurate. The idea is to develop approaches that give good results when the prediction is accurate (consistency) while ensuring that the performance is still acceptable in the worst case, when the prediction is not accurate (robustness). Segment routing is now being widely deployed and used for traffic engineering in IP networks. The key idea in segment routing is to break up the routing path into segments to better control routing paths and improve network utilization. We consider the problem of designing the segments in a network to minimize congestion. This is typically done for a predicted traffic matrix. We use the ideas of consistency and robustness to design a parametrized algorithm that gives good performance when the actual traffic matrix is exactly the predicted traffic matrix (consistency) while giving good performance in the worst case if the actual traffic matrix deviates significantly from the predicted traffic matrix (robustness).