客户端无关的移动TCP性能分析和优化

As the mobile applications (apps) are getting popular, users' requirements for their performance are increasingly high. TCP, the main transport protocol that carries mobile applications, is the dominant factor that impacts the user-perceived performance. However, the high RTT jitter and packet loss rate in mobile and wireless network present huge challenges to TCP performance analysis and optimization for mobile applications. This project aims at extensively analyzing both packet-level and flow-level dataset of several typical mobile services, and tries to tackle two scientific challenges: 1, quantitative analysis of TCP performance bottlenecks and their causes; 2, client-end oblivious TCP performance optimization. To this end, we will study the decision-tree based TCP performance bottleneck diagnosing methodology, which will be further combined with quantitatively analytical method like QED to characterize TCP performance bottlenecks in typical mobile applications. We will also explore the design space of TCP performance improvement for typical mobile applications. Based on the observations, the project will propose a server-end TCP optimization method based on gentle probe packets, as well as a Middle-box based TCP connection optimization method. The proposed performance bottleneck diagnosing methodology and the performance optimization methods will be applied in real-life mobile application platform. The analytical results of TCP performance in typical mobile applications could also shed light on the design of optimal transmission control protocols.

随着移动业务的普及，用户对移动业务的性能要求越来越高。TCP是承载移动业务的主要传输层协议，因此其性能很大程度上决定着用户感知的移动业务性能。无线网络中的RTT抖动问题、高丢包率问题等给移动业务的TCP性能分析和优化提出了巨大的挑战。本项目拟通过对典型移动业务数据包级和业务流级海量数据的分析，围绕TCP性能瓶颈及其原因量化分析和客户端无关的TCP性能优化两个科学问题开展研究。本项目将研究基于决策树的TCP性能瓶颈诊断方法，并结合QED等分析方法，刻画典型移动业务的TCP性能瓶颈，探索性能优化的空间。在此基础上，提出基于探测包的服务器端协议栈优化方法，并研究基于Middlebox的移动网络TCP连接优化方法。本项目提出的TCP性能瓶颈诊断和优化方法预计将在真实移动业务平台部署和应用；而典型业务TCP性能分析的结果将为新型传输控制方法的研究提供科学依据。

TCP是承载移动业务的主要传输层协议，因此其性能很大程度上决定着用户感知的移动业务性能。本项目通过对典型移动业务数据包级和业务流级海量数据的分析，围绕TCP性能瓶颈及其原因量化分析和无客户端参与的TCP性能优化两个科学问题开展了深入研究。提出了基于决策树分类的TCP性能诊断及原因分析方法，并对移动搜索、移动云盘、移动IM等典型移动业务的TCP性能及其瓶颈进行深入分析，探讨了TCP性能提升空间。基于上述结果，设计了多参数感知的TCP拥塞控制机制，包括动态自适应拥塞控制和基于预测的智能丢包检测。最后，研究了基于SDN/NFV的TCP性能优化方法。本项目相关成果在华为广域网存储和奇虎360产品线得到部署应用，华为评价“迅速补齐了我司短板”。在IEEE ToN、JSAC、INFOCOM、IMC等发表学术论文15篇，授权发明专利3项。