2种低延迟服务的通用优化技术  

Two general optimization techniques for low-latency services

作　　者：郭雅宁 徐明伟[1,2] GUO Yaning;XU Mingwei(Institute for Network Sciences and Cyberspace,Tsinghua University,Beijing 100084,China;Zhongguancun National Laboratory,Beijing 100097,China)

机构地区：[1]清华大学网络科学与网络空间研究院,北京100084 [2]中关村国家实验室,北京100097

出　　处：《清华大学学报（自然科学版）》2024年第8期1306-1318,共13页Journal of Tsinghua University(Science and Technology)

摘　　要：近年来,各种低延迟服务吸引了越来越多的用户关注,部署规模不断扩大。为提升低延迟服务性能,工业界和学术界提出多种优化方案并部署在网络传输路径的不同位置,其中部署在端侧的各种低延迟拥塞控制算法和部署在网络侧的主动队列管理算法应用较广泛,这2种算法的设计目标都是尽量避免数据包排队,减少端到端延迟,但是由于这2种算法是独立的,存在潜在的不适配问题,影响应用的性能表现。因此,有关这2种算法的协同优化也成为一个研究方向,基于机器学习的通用算法和端网联合优化是最具代表性的方案。该文总结了低延迟拥塞控制算法和主动队列管理算法的设计思路、组合使用时的性能测试结果以及协同优化的问题,认为跨层联合优化是解决现有不适配问题并进一步提高应用性能的可行思路,建议低延迟服务性能优化的研究应重视通用性和实际部署性。[Significance]Low-latency services have become indispensable in people's work and daily life.Various low-latency services exist,including video conferences for online communication and cloud games for entertainment,which can meet different requirements of users.These services make it convenient for users to interact anywhere in real time,thereby overcoming the limitations of traditional applications.Therefore,these services have recently attracted numerous users,and their deployment scale has rapidly expanded.With the development of 5G technology,the coverage of low-latency services will spread further;these services have broad development prospects.Therefore,performance optimization of low-latency services is a hotspot in academia and industry.The most critical performance indicator for low-latency services is end-to-end latency.In addition to maintaining low latency,achieving high throughput and link usage to improve service quality and attract more users is also necessary.Therefore,performance optimization is key in the further development of low-latency services.[Progress]Various performance optimization schemes used at different positions of the transmission path were proposed by researchers.Among them,the two most widely used schemes were the low-latency congestion control algorithm(CCA)deployed on the server side and the active queue management algorithm(AQM)deployed in the network.Their design tried to avoid queuing as much as possible and to reduce end-to-end delay.The CCA and AQM constantly updated their design to solve the limitations of previous algorithms,improved their performance,and enhanced the practicality of the algorithms for large-scale deployment.Specifically,CCA improved the estimation strategy of congestion signals to make them more accurate,completed the logic of the adjustment of the sending rate and incorporated consideration for fairness into the design.AQM focused on queuing delay and minimized the amount of parameters,trying to implement a more lightweight algorithm.Although CCA and

关 键 词：拥塞控制算法 主动队列管理 低延迟服务 跨层联合优化 机器学习 性能优化 

分 类 号：TP393.0[自动化与计算机技术—计算机应用技术]