【作者】 王圣；

【导师】 苏金树；

【作者基本信息】 国防科学技术大学 ， 计算机科学与技术， 2007， 博士

【摘要】 随着网络技术的快速发展,网络带宽不断得到提升,如何使网络传输更加高效一直是网络研究的热点。目前TCP使用的传输控制机制是针对低速网络而设计,在高速网络环境中面临许多挑战,TCP加速的需求日益迫切。在高速网络TCP加速技术的研究中,至少存在以下问题需要重点解决:如何充分利用路由器信息为端系统的流量控制服务,实现端系统与路由器的协调控制;如何更好地实现流间带宽共享的公平性;如何避免反向数据流量对正向数据流应答的干扰;如何利用高速发展的硬件提高端系统TCP的处理性能等。本文围绕上述关键问题分别从软件和硬件的角度展开深入研究,并提出有效的解决方案。当前端系统与网络之间主要使用ECN进行协调,端系统TCP根据ECN进行速率调整具有较大的盲目性,难以达到拥塞控制的目标。针对这个问题,本文仔细分析了网络资源在TCP传输中的重要性,深入探讨了缓冲振荡现象对TCP性能的影响,提出在报文中引入路由器资源占用情况的指示信息,并提出利用路由器反馈信息进行显式比例带宽分配的方法,同时设计了相应的带宽分配算法EPBA。这种新颖的比例控制方法在端系统与路由器之间建立有效的反馈协调机制,通过比例反馈使分布式调整能够利用更多刚络状态的态势信息,使端系统能够迅速明确地调整发送速率,克服了速率控制的盲目性。同时,EPBA的计算开销很小,易于实现与部署。仿真结果表明,基于路由器辅助的显式比例带宽分配方法能够使端系统获得较好的稳定传输,充分利用了网络资源,从而使TCP性能得到很大的改善。以TCP Vegas为代表的拥塞控制主要采用基于延时的拥塞控制机制,该模型采用固定的速率控制因子,使TCP性能直接与链路的传播延时互成比例,导致带宽分配不均,从而影响全局性能优化。针对这个问题,本文仔细研究了缓冲队列在基于延时拥塞控制机制中的作用,提出了面向全局优化的速率控制因子自适应调整方法,并设计了相应的速率控制因子调整算法RCFAA。在保持基于延时拥塞控制机制原有优越性的前提下,根据链路的传播延时,对速率控制因子进行实时调整,实现流间带宽公平共享,最终达到全局性能优化。仿真结果表明,RCFAA使原有控制机制在公平性上有了较大改善,在性能方面也有较大提升。针对反向数据流量会导致TCP性能下降的问题,本文详细研究了反向数据流量对正向数据流应答的影响,揭示了反向数据流量与TCP传输性能之间的联系,提出了面向单向延时优化的速率控制方法,并设计了相应的速率控制算法UTARC。UTARC将窗口分为更新控制窗口与速率控制窗口,更新控制窗口从底层获取不同类型的延时,并根据这些延时进行相应的窗口调整,而速率控制窗口则控制实际的发送速率。仿真结果表明,在基于延时的拥塞控制机制中,UTARC算法能够有效避免反向数据流量的干扰,充分保证数据链路的利用率,使TCP性能得到较大提升,达到TCP加速的目的。随着网络带宽迅猛发展,协议的处理逐渐成为端系统的沉重负担。针对这个问题,本文对硬件实现TCP的技术展开了详细研究,仔细分析了TCP协议的处理流程,针对TCP流程中的主要操作进行优化的硬件设计。根据硬件特性,提出了面向硬件优化的乱序报文重组方法。最后,作者采用FPGA设计并实现了TCP硬件加速引擎,并对系统中的部分设计优化进行深入探讨。测试结果表明,系统能够大幅度提升TCP处理性能。针对TCP加速的目标,本文从软件和硬件两个角度展开了加速和优化的探索,并取得一些有意义的研究成果。希望本文的研究工作能够有益于今后的TCP加速研究。更多还原

【Abstract】 With the rapid development of network technology, the speed of networks becomes faster and faster. How to make transfer be more efficient becomes the focus of network research. The transmission control mechanism of conventional TCP is designed for previous networks, and faces big challenge in high speed networks. TCP must be accelerated.In the research of TCP acceleration, the issues followed must be considered carefully. First, how to utilize the indication of routers to serve for traffic control in end-systems efficiently, and set up the mechanism for cooperation and control between end-systems and routers. Second, how to obtain better fairness among end-systems for bandwidth allocation. Third, how to avoid the impact of reverse traffic to acknowledgments of forward traffic. Fourth, how to utilize hardware to improve TCP performance in end-systems. We study these critical issues from view of software and hardware, and propose some efficient approaches.Now ECN is used between end-systems and routers. End-systems adjust sending rate according to ECN. Due to the absence of clear network status, end-systems can not adjust sending rate efficiently. We analyze the importance of network resources to TCP performance carefully, and discuss the impact of buffer oscillation. We suggest that some indications of router resources should appear in packet. We also propose an approach of explicit proportional bandwidth allocation with router assistance, called Explicit Proportional Bandwidth Allocation, EPBA, and design an algorithm for it. The novel approach sets up a mechanism for feedback and cooperation between end-systems and routers efficiently, this lets end-systems obtain more indications from routers and adjust sending rate quickly and efficiently with clear objective. What’s more, EPBA can be implemented and deployed easily. Results of our experiments show that EPBA can let end-systems obtain stable transfer, and utilize network resources adequately, and improve TCP performance significantly.TCP Vegas utilizes delay-based congestion control mechanism and uses fixed rate control factor. We find that this lets TCP performance be proportional to link propagation delay and leads unfair bandwidth allocation. We analyze the impact of buffer in delay-based congestion control mechanism carefully, and propose an approach of rate control factor for all-around optimization, called Rate Control Factor Adaptive Adjusting, RCFAA, and design an algorithm for it. RCFAA can not only keep the advantages of delay-based congestion control mechanism, but also adjust rate control factor quickly based on link propagation delay. Results of our experiments show that RCFAA improves fairness in delay-based congestion control significantly and enhances the performance. Reverse traffic can lead degradation of TCP performance. To resolve this problem, we analyze the impact of reverse traffic carefully, and obtain the relation between reverse traffic and TCP performance. We propose an approach of rate control optimized for unidirectional trip time, and design an algorithm for rate control, called Unidirectional Trip Aware Rate Control, UTARC. UTARC splits original congestion window into two parts, update control window and rate control window. The update control window obtains delay from below unit, and adjusts congestion window based on delay. The rate control window is responsible for real sending rate. Results of our experiments show that UTARC can avoid the impact of reverse traffic efficiently in delay-based congestion control, and utilize network resources adequately.With the development of network technology, protocol processing becomes heavy burden of end-system. To alleviate the burden of end-system, we analyze the technology of TCP implementation based on hardware carefully, and study the process of TCP protocol deeply. We optimize the dominating operation of TCP protocol and propose an approach optimized for hardware. This approach is used to reorder disordered packets. At last, we design and implement a system based on FPGA, called TCP Hardware Acceleration Engine, THAE, and study some key issues in THAE deeply. Results of our tests show that our system can improve TCP performance in end-system significantly.To improve TCP performance, we study the technology of TCP acceleration and optimization carefully from view of software and hardware. We hope that our research can help people study TCP acceleration deeply later.更多还原