Scalable offloading using machine learning methods for distributed multi-controller architecture of SDN networks

Machine learning advanced tactics provide flexible assistance to organize, maintain, and optimize SDN flat topology, despite that intelligence is hard to apply and deployed in the SDN domain. Knowledge-defined network modeling opens a new challengeable door to build a self-driving SDN framework with better precision and efficient performance, all of this encourages the proposal of a novel EQUI-dispatcher model for flat topology. The model is designed on the advanced learning capability of gated graph recurrent neural networks (GG-Rec-NNs). Our proposed model has a generalized capability on variable traffic load and delays under disparate routing schemes. In GG-Rec-NNs several learnable parameters are liberated size and the sequence's length over arbitrary topologies. The graph illustrates the secular structure of sequence index and spatial structure support, which guaranteed scalability. Recurrent neural network (RNN) is used to capture hidden state dependencies on a sequence of variable size for link-level message aggregation, along with this vanishing gradient problem come across. To overcome this problem, gate (on Link and Path) is embedded to encode long-range dependencies, during training traffic data and routing not visible when a test against topologies. We also present model potential in the light of numerical experiments or results based on real and synthetic datasets that support its feasibility as compared to traditional routing strategies for SDN networks.