CNN-based Methods of Traffic Matrix Prediction with Multiple Timesteps in Optical Network-on-Chip (ONoC)

Optical Network-on-Chip (ONoC) is a promising type of interconnection with advantages such as high bandwidth and low latency. The network performance of ONoC can be further enhanced by accurate traffic matrix prediction. Convolutional neural network (CNN) can capture the spatial-temporal information of the input through convolution kernels. In this paper, three CNN-based methods (i.e., Multi-Width CNN, Multi-Channel CNN, and Multi-Input CNN) of traffic matrix prediction are proposed, with multiple timesteps. Multi-Width CNN and Multi-Channel CNN are Single-Input CNN. The evaluation results show that the traffic prediction accuracy of the Multi-Input CNN is significantly improved, compared to the single input. In addition, the Multi-Input CNN better improves performance of ONoC, compared to some popular neural network-based prediction methods.