Congestion Management In The Highly Dynamic Network

The Internet has become an indispensable infrastructure across the world and the network techniques draw a growing attention from the academia and the industry.To meet the various requirements of different network applica-tions,novel network techniques,e.g.,novel network hardware,novel network application deployment model and novel network architecture,are introduced and integrated with current network.However,the continuous integration of new network techniques also brings intractable network dynamics and causes the network management problem.Moreover,the management problem comes along with the big surge of network traffic and all the problems pose a great challenge for current traffic control.Therefore,it is critical to provide pre-dictable transmission performance in the highly dynamic network and control the network congestion efficiently.This dissertation studies the network con-gestion management in the highly dynamic network from different perspectives and the contributions are as follows:(1)The objective-oriented Multi-path TCP(MPTCP)congestion control algorithm.The traditional MPTCP congestion control algorithms are ag-nostic to the objectives of network applications,which incurs the lack of the capability to adapt and optimize for different applications.To over-come this problem,this dissertation designs a novel deep reinforcement learning(DRL)-based MPTCP congestion control algorithm,Partner.By designing different reward functions,Partner can use a universal frame-work to optimize for different objectives and maximize their performance.The simulation results show that Partner can utilize a universal frame-work to cooperate' with different packet schedulers and satisfy different applications.(2)The MPTCP congestion control algorithm via imitation learning.The casual usage of DRL in MPTCP congestion control algorithms may re-sult in serious unfairness and unpredictable performance.To solve the problem,this dissertation proposes a novel MPTCP congestion control algorithm via imitation learning,IMCC.IMCC also combines the imita-tion learning-based congestion control algorithm with a backup congestion control algorithm to provide consistent performance.In the familiar en-vironment,the imitation learning-based algorithm will take effect,while the backup algorithm will handle the unseen environment.The simulation results show that IMCC successfully handles the high dynamics in MPTCP and achieves stable performance and fairness.(3)The core-stateless network performance isolation in the public cloud.Current works cannot provide the high scalability and high performance simultaneously.To handle this dilemma,this dissertation designs a core-stateless network performance isolation method,SLIT.SLIT moves the scheduling intelligence to the network edge(hypervisor)and makes the network switches schedule queuing packets based on the information pack-ets carry.As a result,SLIT successfully mimics the scheduling behavior of the ideal switch with the multiple physical queues.The simulation and testbed results show that SLIT accomplishes good bandwidth isolation across virtual machines(VMs).Moreover,it also provides fast conver-gence,bandwidth utilization improvement and short flow friendliness.(4)The congestion management with ultra-low latency data plane in the process of the software-defined networking(SDN)update.Because of the long control loop introduced in current SDN update methods,they cannot adapt to the rapid traffic fluctuation in the emerging ultra-low latency data plane.To overcome this problem,this dissertation proposes an SDN update method with flow rate estimation in the data plane,MDVP.It considers the traffic change in the process of network update and estimates the converged flow rate in each network state.As a result,it avoids the uncontrolled congestion caused by the dramatic traffic change in the data plane.The simulation results show that MDVP can make a flexible trade-off between the number of intermediate states and the extent of network congestion.Besides,it can reduce the demand violation ratio significantly compared with the state of the arts.