Multi-Step Forecasting of Intense Traffic Streams Using Machine Learning for Optical Circuit Switching

In this work, we investigate the performance of recurrent nonlinear autoregressive neural networks for an efficient prediction of intense data traffic streams for application in hybrid electro-optical intra-data center networks. The main motivation behind this approach is to evaluate the traffic patterns generated by servers hosting different services and to forecast multiple steps ahead the occurrence of heavy traffic streams using machine learning. This allows an a priori scheduling and allocation of optical switching circuits for offloading of intense data traffic from electrical packet switches. In this respect, for a realistic analysis, real server generated data gathered over a two-week time span have been used in our simulative evaluation. As with the synthetic data sets used in our previous work, the recurrent nonlinear autoregressive neural networks show a promising performance of the multi-step forecasting procedure.