GCLR: GNN-Based Cross Layer Optimization for Multipath TCP by Routing

Multipath TCP has attracted increasing attention as a promising technology for 5G networks. To fully utilize network interfaces on multi-homed terminals and the whole network resources, MPTCP is proposed as an extension of TCP to transfer packets concurrently over multiple paths. Cross layer optimization techniques have been applied for MPTCP such as routing and path management. However, existing multipath routing algorithms and network modeling techniques are facing the challenges of subflow asymmetry due to network heterogeneity, thus cannot handle routing optimization problems comprehensively. To address these problems, in this paper, firstly, a novel Graph Neural Network (GNN) based multipath routing model is proposed to explore the complications among links, paths, subflows and the MPTCP connection on various topologies. Leveraging the GNN model, expected throughput can be predicted with given network topology and multipath routes, which can further be the guidance for optimzing the multipath routing. Then, GCLR, a GNN based cross layer optimization system for MPTCP by routing, is proposed with the help of SDN (Software Defined Networking). According to simulation results, our off-line learned GNN model can predict the expected throughput of specific MPTCP connections with very low error. Besides, it's validated that the model has high generalization ability in terms of connection arbitrary and topology arbitrary, it can maintain MSE (mean squared error) at a low level when the situations are not seen during training, which is sufficient for throughput prediction in multipath routing decisions. Finally, the online routing optimization system is realized using SDN, experimental results show that our proposed routing optimization system can achieve significant throughput enhancement compared with traditional multipath routing algorithms.