Treffer: Data center traffic scheduling algorithm based on spatial–temporal graph convolution networks.

To address the problems of time delay in traffic scheduling and large traffic collision in data center network, a link load balancing strategy based on the spatial–temporal graph convolution networks is proposed. Firstly, a link load state prediction model is designed. Based on the topologic structure and traffic characteristics of data center network, this thesis uses graph convolutional neural network (GCN) and recurrent neural network (RNN), to extract the spatial characteristics of data center network and the time-order characteristics of traffic, thereby improved the model's perception precision of link load. Then, a link load balancing algorithm based on spatial–temporal graph convolution networks is proposed, the algorithm measures the link selectivity according to the predicted information of link load state and the actual link information obtained by the controller. With the objective to minimize the collision of large traffic at the core layer, uniform distribution of traffic can be achieved. Next, by combining the comprehensive index of link, the selectivity of path can be determined. An improved artificial bee colony algorithm is introduced to calculate the routing strategy of large flow, and single-dimensional neighborhood search is optimized to multi-dimensional neighborhood search to accelerate the convergence speed. The simulation results show that the proposed load balancing strategy can effectively enhance the load balancing, reduce the transmission delay, and improve the bisection width of the network. [ABSTRACT FROM AUTHOR]