【作者】 李琦；

【导师】 郑烇；

【作者基本信息】 中国科学技术大学 ， 网络空间安全， 2023， 硕士

【摘要】 随着互联网的快速发展,网络中承载的流量越来越多,拥塞控制算法的作用日益凸显。良好的拥塞控制算法是数据高效稳定传输的重要保障。近年来在网络空间安全领域中,数据异地容灾备份、网络安全态势感知、网络威胁情报共享等热点研究内容均涉及到大数据传输。大数据传输场景在收敛性能、重传和适应性等方面对拥塞控制算法提出了更高的要求,而已有的拥塞控制算法无法满足。目前性能较优的 BBR(Bottleneck Bandwidth and Round-trip propagation time)算法及其衍生算法在大数据传输场景下收敛性能和重传表现欠佳,而引入细粒度网络信息反馈的算法不适合长距离传输,且往往依赖专用硬件,增大了部署的难度和成本。针对上述问题,本文分别从端系统和端网协同两个方向进行优化。在端侧,本文建立了 BBR流在探测带宽阶段的收敛行为模型,提出了一种考虑网络缓存资源的拥塞控制算法;同时在大数据传输场景中引入了细粒度网络反馈信息,提出了一种事件驱动的端网协同拥塞控制算法。具体的研究工作如下:(1)本文提出了一种基于BBR的增益自适应拥塞控制算法。基于BBR及其衍生算法没有考虑网络缓存资源的问题,提出了一种收敛域模型,描述了 BBR算法在探测带宽阶段各个流的收敛行为。然后将各个BBR流快速收敛且不超出缓冲区的问题转化为优化问题,求解后得到发送速率增益的初始取值策略。随后进一步考虑网络拥塞情况,在发送速率增益的取值策略中引入调节因子,进一步降低发送速率增益来减缓拥塞。(2)本文提出了一种基于事件驱动的端网协同拥塞控制算法。将细粒度的网络信息反馈引入到大数据传输场景下的拥塞控制的同时,降低了对网络设备的要求。在网络设备上实现了网络状态信息采集和事件触发上报机制,在端系统上实现了网络拥塞程度划分和分阶段拥塞控制策略。考虑到网络的动态变化特性,设计了 一种基于缓存梯度的窗口调节策略进行窗口控制,结合缓存占比和缓存占用变化,预测后续的缓存占用大小,从而最大程度利用带宽资源。本文进行了大量的实验,通过与基准算法进行性能对比分析,证明了基于BBR的增益自适应拥塞控制算法和基于事件驱动的端网协同拥塞控制算法具有优越的性能,在大数据传输场景下实现了更高的传输速度、更低的重传和更快的收敛速度。更多还原

【Abstract】 With the rapid development of the Internet,more and more traffic is carried in the network,and the role of congestion control algorithms is becoming more and more prominent.A good congestion control algorithm is an important guarantee for efficient and stable data transmission.In recent years in the field of cyberspace security,hot research contents such as data off-site disaster recovery and backup,network security situational awareness,and network threat intelligence sharing are involved in big data transmission.The big data transmission scenarios put forward higher requirements on congestion control algorithms in terms of convergence performance,retransmission and adaptability,which cannot be met by existing congestion control algorithms.The current BBR(Bottleneck Bandwidth and Round-trip propagation time)algorithm and its derivatives have poor convergence performance and retransmission in big data transmission scenarios,while the algorithms that introduce fine-grained network information feedback are not suitable for long-distance transmission and often rely on dedicated hardware,which increases the cost and difficulty of deployment.To address the above problems,this thesis optimizes the end-system and endnetwork collaboration in two directions,respectively.At the end side,this thesis establishes a convergence behavior model of BBR flows in the probing bandwidth stage and proposes a congestion control algorithm considering network cache resources;meanwhile,it introduces fine-grained network feedback information in the big data transmission scenario and proposes an event-driven end-network cooperative congestion control algorithm.The specific research work is as follows:(1)In this thesis,a BBR-based gain-adaptive congestion control algorithm is proposed.Based on the problem that BBR and its derived algorithms do not consider network cache resources,a convergence domain model is proposed to describe the convergence behavior of each flow of the BBR algorithm in the probing bandwidth phase.Then the problem of fast convergence of each BBR flow without exceeding the buffer is transformed into an optimization problem,and the initial value strategy for the transmit rate gain is obtained after solving.Then,the network congestion is further considered,and a modulation factor is introduced into the sending rate gain acquisition strategy to further reduce the sending rate gain to mitigate the congestion.(2)This thesis proposes an event-driven end-network cooperative congestion control algorithm.The fine-grained network information feedback is introduced to the congestion control in the big data transmission scenario while reducing the requirements on the network equipment.The network state information collection and event-triggered reporting mechanism are implemented on the network devices,and the network congestion degree division and phased congestion control strategy are implemented on the end system.Considering the dynamic change characteristics of the network,a window adjustment strategy based on cache gradient is designed for window control,which combines the cache occupancy ratio and cache occupancy change to predict the subsequent cache occupancy size,so as to maximize the utilization of bandwidth resources.In this thesis,we have conducted extensive experiments and demonstrated the superior performance of the BBR-based gain adaptive congestion control algorithm and the event-driven end-network cooperative congestion control algorithm by comparing performance analysis with benchmark algorithms to achieve higher transmission speed,lower retransmission and faster convergence in large data transmission scenarios.更多还原