软件定义广域网络中基于带内探测的流量调度机制

In recent years, with the popularity and booming of Internet services such as web search and 4K/8K video, the wide area network has become an important factor affecting the user experience. At present, the traditional wide area network leverages the distributed mechanism to manage a large number of switches and routers. In order to make full use of network resources, the software-defined WAN and the traffic scheduling mechanisms have received extensive attention from industry and academia. However, due to the extremely complex topology of the WAN and the variegated requirements of network services, the traditional traffic scheduling algorithms encounter performance bottlenecks and many problems in the wide area network: 1) It is difficult to quickly measure the delay and bandwidth for large-scale networks 2) Imbalanced traffic schedule usually causes network congestions in the multipath topology 3) The delay requirement of delay-sensitivity service is difficult to be guaranteed. In this project, we first propose an in-band measurement mechanism based on active probing. Then, we propose a distributed load balancing mechanism based on in-band measurement for SD-WAN, and use a random algorithm to avoid the herd behavior problem. Finally, we propose an QoS algorithm for multi-objective for reducing the latency of latency-sensitive traffic.

近年来，随着网络搜索、4K/8K视频等互联网业务的普及与蓬勃发展，支撑这些服务的广域网络成为影响用户体验的一个重要因素。当前传统广域网采用分布式机制管理大量交换路由设备，为了充分利用网络资源，软件定义广域网和其中的流量调度机制得到了工业界和学术界广泛关注。然而，由于广域网具有极其复杂的拓扑结构并包含了需求各异的业务，导致传统的流量调度算法在广域网络中遇到了性能瓶颈和许多不同的问题：1）大规模网络下路径时延与带宽难以快速测量；2）多径拓扑下由流量调度不均引起的路径拥塞问题；3）多业务场景下不同时延敏感程度业务流的时延需求难以得到保障问题。本课题首先研究基于主动探测的低开销带内测量机制；然后提出基于带内测量的分布式路径选择机制，并引入随机性来避免羊群效应问题；最后，提出多目标结合的服务质量保障机制，满足不同业务对带宽、时延等的需求，减小时延敏感业务完成时间。

面向4K/8K视频、实时操作等新兴互联网应用，广域网亟待变革。针对大规模网络下路径时延与带宽难以快速测量、多径拓扑下由流量调度不均引起的路径拥塞问题，本项目基于大规模快速流估算（LFE）框架，提出了面向混合流调度的智能流长预测机制，与FLUX相比，LFE将平均FCT提升了1.01-1.13倍；提出信息无感知协同流调度机制DeepAalo，相比于Aalo机制，DeepAalo将Coflow的平均完成时间提升了高达1.37倍；提出了一种面向多队列的自适应拥塞标记调节机制A-ECN，与DCTCP的ECN标记方案相比，A-ECN优化短流的流完成时间达40%。针对多业务场景下不同时延敏感程度业务流的时延需求保障问题，本项目设计了面向大规模骨干网的确定性边缘整形机制、多类型混合流共传机制、以及在线时延带宽等多目标结合的流量调度算法，通过边缘整形，网络资源利用率提升了20%以上，且与传统离线计算相比，在线调度的计算时间从分钟级缩小到秒级/毫秒级。本项目突破了基于软件定义的广域确定性骨干网系统架构、解决了长距链路精确转发、服务质量有界保障等难题；依托国家未来网络试验设施CENI，完成了全球首个确定性广域网的创新技术试验，构建了覆盖北京、南京、上海等13个城市核心节点的大规模试验网，首次成功试验了跨2000公里以上传输距离、小于30微秒级的时延抖动精确控制，并开展了工业机器远程控制业务。