【作者】 张龙；

【导师】 王晶；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 在5G网络中,僵化的TCP不能满足实时的数据传输要求,新兴的 QUIC(Quick UDP Internet Communication)协议有望成为 TCP的替代方案。同时,多路径技术可以提高数据传输的有效性和可靠性,然而,传统的调度器无法适应复杂和动态的5G网络。因此,当多路径条件实时变化时,如何制定路径选择策略和多流调度算法,以减轻流内阻塞和流间阻塞问题,是在实际的5G场景下非常值得探讨的问题。本文针对路径选择和多流调度过程,在5G动态网络场景下研究了多路径调度策略,并从减轻接收端的流内阻塞和流间阻塞这两个不同的角度提出了解决方案。一、为了缓解多路径选择导致的流内阻塞问题,本文基于MPQUIC(multi-path QUIC)协议,提出了一种基于深度强化学习的路径选择策略DeepPath。DeepPath采用了一种定制的双DQN(Deep Q-Network)算法来执行最优的策略。为了有效地对5G网络的动态性和异质性建模,DeepPath在状态的设置中充分考虑了往返时延(Round-Trip Time,RTT)的平均值、RTT的标准差、拥塞窗口的大小和平均排队长度。离线部分通过双网络联合训练以精确地拟合路径价值。在线部分,多路径调度器实时进行贪心决策以获得最大的吞吐量。DeepPath在仿真和真实的5G场景下进行了广泛的测试,结果表明优于现有的多路径调度器。二、为了缓解多流之间的阻塞问题,本文基于流优先级提出了多流组包算法PriorityStream。PriorityStream共包含两个策略。策略一旨在保障每个优先级的流不会被阻塞,在一个调度周期中,根据优先级的占比确定各个流发包的数量,并让高优先级的流优先发包。策略二为了解决传统的重注入算法存在的阻塞问题,首先,检查未确认队列中是否有大于等于高优先级流的数据包。然后,当调度完高优先级流的最后一个数据包后,立刻检查未确认队列中相同优先级的数据包,以避免低优先级的流阻塞高优先级的流。实验证明,本文提出的多路径调度器(DeepPath+PriorityStream)在性能上优于现有的调度算法。本文还搭建了一个完整的5G实验系统,包括一个5G核心网、一个5G终端和一个基站。整个系统支持多用户同时注册、建立协议数据单元会话和访问数据网络等功能。实验在该5G核心网中开展,测试上述方案的有效性和网络吞吐量。结果表明,通过改进的MPQUIC协议来替换HTTP/2,减少了各网元间信令的传输时间,从而显著提升了信令传输效率。更多还原

【Abstract】 In 5G networks,it has been observed that traditional Transmission Control Protocol(TCP)is insufficient to meet the demands of real-time data transmission.As an alternative,Quick UDP Internet Communication(QUIC)has emerged as a promising solution.Additionally,multi-path technology has been proposed to enhance the efficiency and reliability of application data transmission.However,conventional schedulers face challenges in adapting to the complex and dynamic 5G networks.Consequently,addressing real-time changes in multi-path conditions and devising effective packet scheduling strategies and multi-stream management algorithms to mitigate intra-stream and inter-stream Head-ofLine(HoL)blocking issues respectively has become a hot topic.With a focus on path selection and multi-stream scheduling,this study aims to investigate multi-path algorithms in dynamic 5G network scenarios.Specifically,solutions are proposed from two perspectives:alleviating intra-stream and inter-stream HoL blocking issues at the receiving end.Firstly,to address the issue of intra-stream Head-of-Line(HoL)blocking in multi-path packet scheduling,we propose a Deep Reinforcement Learning(DRL)based path selection strategy,called DeepPath,for the multi-path Quick UDP Internet Communication(MPQUIC)protocol.DeepPath employs a customized Deep Q-Network(DQN)algorithm to learn an optimal policy.To accurately model the dynamic and heterogeneous nature of 5G networks,DeepPath takes various factors into consideration,such as Round-Trip Time(RTT),standard deviation of RTT,size of congestion window,and average length of queue,as part of the state setting.The double network is jointly trained to accurately approximate the path value in the offline phase.In the online phase,the multi-path scheduler makes real-time greedy decisions to achieve maximum goodput.DeepPath has been extensively evaluated in both simulated and real-world 5G scenarios,demonstrating superior performance compared to other multi-path schedulers.Secondly,to mitigate the inter-stream Head-of-Line(HoL)blocking issue,we propose a stream priority based multi-stream management algorithm in this paper,termed PriorityStream.PriorityStream encompasses two strategies.In the first strategy,the number of packets sent in a scheduling period is determined based on the proportion of priorities,ensuring that packets of higher priority are transmitted first to prevent blocking of streams at each priority level.In the second strategy,to overcome the blocking issue associated with traditional reinjection algorithms,we firstly check whether there are packets in the unacknowledged queue that are greater than or equal to the high-priority stream.Then we introduce an immediate checking mechanism for packets of the same priority in the unacknowledged queue when the last packet of a higher priority stream is sent.This approach effectively prevents low priority streams from blocking high priority streams.Through experimental evaluations,we demonstrate that our scheduler,consisting of DeepPath for path selection and PriorityStream for multi-stream management,outperforms other scheduling algorithms.We have also established a complete 5G network simulation environment,consisting of a 5G core network,a terminal,and a base station.We conduct experiments in the 5G core network to evaluate the effectiveness of the proposed scheme and the network throughput.The results show that by replacing HTTP/2 with our improved MPQUIC protocol,the time required for signaling transmission between network functions is reduced,leading to significant improvement in signaling transmission efficiency.更多还原