Flowlet-level multipath routing based on graph neural network in OpenFlow-based SDN

Abstract

Software-Defined Networking (SDN) revolutionizes the rigid operation and management mode of traditional vertical architecture. A centralized controller enables more granular control of the network based on a global view. However, the traffic forwarding mode in the traditional networks is still inherited, which cannot make full use of the underlying network resources to meet higher quality of service. In addition, the demand for massive data processing has also forced operators to seek the help of intelligent management solutions. In this paper, we propose a flowlet-level multipath routing scheme based on graph neural network to leverage the performance optimization benefits of parallel transport. The flowlet-based solution strikes a balance between flow transmission granularity and reordering, which can effectively improve the end-to-end transmission efficiency. Specifically, first, we design an adaptive flow splitting scheme to achieve multipath transmission when the network state changes dynamically by using the rule timeout mechanism that comes with the OpenFlow protocol. Secondly, we propose to use graph neural network to predict link delay to assist in adaptive flow splitting and forwarding path selection intelligently. Simulation results show that the graph neural network has good convergence and generalization in delay prediction, and the flow splitting mechanism based on rule timeout realizes adaptive flow splitting according to network state changes. The overall scheme outperforms existing typical solutions in terms of time overhead, end-to-end delay, flow completion time, and throughput. Meantime, experiments on different topologies verify the scalability of our scheme.