Efficient Network Traffic Prediction after a Node Failure

Currently, we observe a high popularity of the traffic-aware network management and optimization approaches, which benefit from the traffic modeling and prediction tools. The efficiency of these approaches depends on the accuracy of the applied modeling and prediction methods, which might be significantly decreased by exceptional events and anomalies, like for instance a long-lasting node failure. After such cases, the modeling and prediction tools may provide low-accuracy and misleading data, which used as an input to management/optimization methods might significantly decrease the network performance. Therefore, it is crucial to evaluate the approaches after such events, draw conclusions regarding their reliability and define application instructions for some special cases. The presented study answers that problem and evaluates how much we can rely on the traffic modeling and prediction approaches when a node failure occurs in a network. It compares a number of approaches and tries to select the most reliable one. The main comparison criterion relates to the time necessary to detect a change in the traffic pattern, adapt models to that event and restore a system convergence.