Long-Term Traffic Scheduling Based on Stacked Bidirectional Recurrent Neural Networks in Inter-Datacenter Optical Networks

Abstract

With the rapid evolution of high-speed mobile communications, cloud computing, and other high-bitrate datacenter-supported services, efficient and flexible traffic scheduling has become one of the fundamental tasks of inter-datacenter optical networks (IDCONs). Traffic scheduling algorithms based on long-term traffic prediction, which have intelligent and global resource allocation ability, have been proved to perform well in IDCONs. However, the low accuracy of existing long-term traffic prediction methods, which is caused by the accumulated errors produced in the recursive multi-step prediction process, directly restricts the efficiency of traffic scheduling. In this paper, we consider the problem of highly efficient traffic scheduling in IDCONs by leveraging one step long-term traffic prediction to reduce the prediction errors. We first design a multiple time interval feature-learning network (MTIFLN) to handle the challenging task of one step long-term traffic prediction. By integrating five bidirectional RNNs (B-RNNs) to one single framework, the MTIFLN has a strong ability to extract the long-term traffic features at different time intervals. Moreover, the stacked architecture of MTIFLN helps to reduce the prediction errors through multi-resampling process. A traffic prediction-based resource allocation (TP-RA) algorithm is proposed together with a global factor to evaluate the efficiency of traffic prediction and achieve effective traffic scheduling based on both traffic prediction results and network resource utilization. Simulation results indicate that with our proposal, the MTIFLN can accurately predict the traffic for more than 24 hours in one step, and the TP-RA algorithm enables IDCONs to make more efficient use of network resources.