Graph Neural Network Based Performance Optimization For MPTCP In HetNet

Multipath Transport Control Protocol(MPTCP)is the extenstion of Transmission Control Protocol(TCP)for multipath parallel transmission.Compared with traditional TCP,MPTCP has the advantages of high throughput,high robustness,and easy load balancing.The proposal of MPTCP has attracted widespread attention from academia to industry,and has produced a series of applications from mobile communication net-works to data center networks.Due to different network interfaces equipped by multi-homed hosts,asymmetric link characteristics and different subflow paths lead to the difference of subflow trans-mission capacity,making it dificult for MPTCP connections to achieve ideal perfor-mances in heterogeneous networks.To solve this problem,first of all,the multipath routing algorithm can reduce subflow asymmetry from the perspective of network layer.Secondly,in the case of subflow asymmetry,transmission control algorithms such as packet scheduling and congestion control performs differentiated resource allocation ac-cording to the state of subflows,and alleviates the performance degradation of MPTCP connections in heterogeneous networks from different perspectives.Existing solutions are difficult to take advantage of the multiple paths brought by MPTCP in heterogeneous networks.For example,traditional heuristic algorithms make ideal assumptions for a specific problem,and propose simple algorithms based on sim-plified models.Therefore,they can only achieve the expected performance in specific scenarios,and are difficult to cope with complex real scenarios.Although the emerg-ing machine learning-based algorithms have stronger expressive ability and are suitable for more complex scenarios,they can only take effect in training scenarios.In online applications,they rely on real-time updates of the model,which brings additional com-putational overhead.Graph Neural Network(GNN)is a new type of neural network model.Through message passing among nodes,it analyzes graph-structured data,and learns the rela-tionship among nodes and the structural characteristics of graph.It has been proved that GNN has remarkable expression ability and generalization ability.In the transmission control of MPTCP,the relationship among different transmission control decisions,link characteristics of subflows,and connection performance can be represented by graph.Therefore,in this dissertation we propose GNN based MPTCP performance optimiza-tion in heterogeneous networks,from the perspectives of multipath routing,MPTCP packet scheduling and congestion control,the following algorithms are proposed(1)GNN based multipath routing using SDN(Software Defined Network).To alleviate the impact of subflow asymmetry on the performance of MPTCP connections,a multipath routing algorithm is proposed with the function of controling the number of subflows,avoiding sharing bottleneck links,and matching the capacity of subflows.The system obtains the network topology information through SDN,trains a GNN based throughput prediction model for multipath routing schemes,and learns the relationship between connection performance and subflow link characteristics.Finally,in the online applications,the optimal routing strategy is determined according to the throughput prediction results of different candidate routing schemes.Thanks to the ability of GNN to infer new graphs,the algorithm performances better than default multipath routing algorithms even facing unknown scenarios that have never appeared in training.(2)GNN based MPTCP Scheduler using SDN.Aiming at the phenomenon that existing algorithms cannot achieve ideal performance in heterogeneous networks,the algorithm optimizes MPTCP performance from different stages of the MPTCP connec-tion:when the number of subflows changes,the subflow management module disables the subflows with poor performance,when new packets are to be sent,the data schedul-ing module selects an appropriate subflow for scheduling.Based on the relationship among link information,subflow information,and connection information in MPTCP scheduling,we propose a GNN-based throughput prediction model for subflow manage-ment and data scheduling schemes,then do subflow management and packet scheduling separately according to the guidance of the model prediction results.The algorithm can improve the performance of MPTCP connections in heterogeneous networks,and has the generalization ability in terms of connection characteristics and topology.(3)GNN based MPTCP Congestion Control.Aiming at the shortcomings of con-gestion judgment and window adjustment of existing algorithms,a congestion control algorithm is proposed to ensure network fairness.We use GNN model to learn the rela-tionship between network fairness and MPTCP subflow state,and evaluate the impact of different window decisions on network fairness in each congestion control round,and then adjust the window.The algorithm can ensure network fairness of MPTCP connec-tions in heterogeneous networks,and improves the performance of MPTCP connections in packet loss environment.