【作者】 孙伟；

【导师】 温涛；

【作者基本信息】 东北大学 ， 计算机应用技术， 2010， 博士

【摘要】 流媒体技术是宽带通信网络和多媒体技术共同发展的产物。近年来,随着互联网、通信技术、多媒体压缩技术以及终端处理能力的快速发展,互联网流媒体技术应运而生。然而由于Internet与生俱来的“尽力而为”的服务特性以及Internet为数据传输而设计的初衷,加之流媒体应用的特殊服务质量要求,大范围地在互联网上部署流媒体应用仍然面临许多挑战。这其中,如何在高效地传递流媒体数据的同时保证互联网的稳定性是急需解决的问题。针对以上问题,目前国内外学者已提出多种流媒体传输速率及拥塞控制机制。在众多的解决方案中,目前被公认为最有效的流媒体传输速率控制方案是TFRC (TCP Friendly Rate Control, TCP友好性速率控制)协议。IETF也制定了一系列TFRC相关的RFC文档。然而,大量的研究表明,TFRC协议仍然存在一些缺陷。首先,TFRC采用的基于TCP Reno的TCP吞吐量方程并不能反映目前在互联网中占主流地位的TCP NewReno数据流的吞吐量。其次,TFRC采用了类TCP协议的慢启动算法,很容易导致在慢启动结束时大量的数据包丢失。第二,TFRC丢包判定算法受到网络中乱序数据包的影响,容易发生丢包误判。本文针对以上问题展开研究,并提出了相应的解决方案。所取得的具体成果如下：(1)提出了一种基于TCP NewReno的TCP稳态吞吐量分析模型,用于描述TCP数据流的吞吐量与往返时延,重传超时时间以及丢包事件概率之间的关系。该模型充分考虑了慢启动与快速重传阶段对TCP吞吐量的影响。仿真实验表明,该模型可以准确地预测网络中TCP NewReno数据流的吞吐量。将该模型应用于TFRC协议,可以很好地保证TFRC协议对于TCP协议的友好性。(2)提出了一种基于带宽测量技术的TFRC慢启动算法。该算法利用在线网络带宽测量技术,探测出目前网络可用带宽,从而根据网络状态实现发送速率的动态更新。同时改变发送速率的增长方式,使发送速率在连接启动时增幅较大,而在慢启动结束过度到拥塞避免阶段增加幅度较小。仿真实验表明,该算法有效避免了多个分组丢失现象的发生,提高了TFRC连接由慢启动阶段过渡到拥塞避免阶段的平滑性。(3)提出了一种基于延时响应的丢包判定算法,用于克服乱序数据包对TFRC协议的影响。该算法在接收端收到3个乱序数据包时会启动一个延时定时器,以允许缺失的数据包通过其他路径或链路层的恢复机制到达接收端。仿真实验表明,该算法可以有效降低乱序数据包出现的次数,降低丢包的误判率。而当网络中无乱序数据包时,改进算法仍对标准TCP协议具有很好的公平性和友好性。(4)分别在NS2仿真实验平台和Linux操作系统中实现了改进的TFRC协议。并通过大量的仿真实验与真实网络测量考察了改进协议的性能。结果表明改进协议在传输速率平滑性,对于TCP协议的友好性,对拥塞响应的稳定性等多方面均可以满足流媒体应用对于传输速率控制的要求。更多还原

【Abstract】 Streaming media is a combination of communication and multimedia technology. In recent years, with the rapid development of the Internet, workstation performance, communication technologies and multimedia compression technologies, it is possible to provide streaming media applications over the Internet. However, due to the best-effort property of the Internet and features of streaming applications, large-scale deployment of streaming media applications is still facing many challenges. In particular, how to effectively deliver the data produced by such streaming media is the key problem for streaming media applications.There has been a significant amount of previous work on the transmission rate control mechanism for streaming media and several rate control protocols have been developed. Among them, the TFRC (TCP Friendly Rate Control) protocol is the most promising solution. It has gained much popularity as a reference scheme for streaming media transport on the Internet. In fact, TFRC has become a standard RFC in 2004.However, On the basis of the systemic analysis and performance evaluation of TFRC protocol, we found that for safety and large scale deployment of TFRC on the Internet there are still some problems to be resolved. First, the TCP throughput equation used by TFRC was based on TCP Reno, which can not accurately predict the throughput of TCP NewReno, the main implementation of TCP protocol on today’s Internet. Second, TFRC’s TCP-like slow start algorithm may result in a large number of packet losses within one round trip time, which degrade the performance of TFRC and impact the quality perceived by the receiver. Third, TFRC gives poor performance over network scenarios with packet reordering because it emulates TCP and treats a packet reordered beyond 3 as lost packet.The research presented in this dissertation is concentrated on improving the performance of TFRC protocol. Different schemes of improving the performance of TFRC protocol is proposed respectively at TCP throughput equation, slow start algorithm and packet loss detection algorithm. The main contributions of this dissertation are as follows:First of all, this dissertation develops a simple and accurate analytic model for the steady state throughput of the slow but steady variant of TCP NewReno by capturing the effect of fast recovery algorithm and taking into consideration slow start phase after timeout expiration. The model describes the relation between TCP NewReno throughput and round trip time, loss event rate and retransmission timeout value. Validation by NS2 simulator shows that using TFRC’s throughput model to estimate TCP NewReno throughput may introduce significant error and the proposed model is able to accurately predict the steady-state throughput for TCP NewReno over a wide range of network conditions.Secondly, a bandwidth measurement based slow start algorithm was proposed in this dissertation. The algorithm employs effective online bandwidth measurement technology to get the available bandwidth and update the sending rate with appropriate value dynamically. It increases the sending rate with half of the sum of the current sending rate and the measured bandwidth, iterates and gradually closes up available bandwidth, of which the increment is large at start phase and small at end phase of a connection. Simulation experiments indicate that the algorithm significantly decreases the dropped packets and improves the smoothness of connections.Thirdly, this dissertation presents a delay based packet loss detection algorithm to make TFRC more robust to packet reordering and yet, when packet reordering does not occur, it is friendly to the standard implementation of TCP. In TFRC, the loss of a packet is detected by the arrival of three packets with higher sequence number. This packet loss decision is delayed in the new algorithm by a short period to allow the receiver to receive the packets that travel in different path or the link level mechanism to recover the lost packet. If at the end of the delay timer the packet is still not received, then it is treated as a packet loss due to congestion. The simulation results show that the algorithm performs consistently better than the standard TFRC under persistent packet reordering. When the case that packets are not reordered it maintains the same throughput as the typical implementation of TCP (TCP-NewReno) and shares network resource fairly.Finally, the refined TFRC protocol was implemented in NS2 simulation environment and Linux operation system. And we extensively evaluated the refined TFRC protocol through simulation and real network measurement. The results show that it is a suitable rate control mechanism for streaming media in the aspects of smoothness, TCP friendliness and the responsiveness of congestion.更多还原