Study And Implementation Of High Performance Network Congestion Control Protocol And Key Algorithm

Congestion control is a key technique in the field of computer networks. The early network congestion control has focused on improving TCP congestion control and got some achievements. But with the use of fiber,and the emerage of high performance router and a large amount of application software, the network's performance is improved very quickly and presents a character of large scale, high speed, large bandwidth, long delay and more and more kinds of applications. So the traditional TCP congestion control and the drop tail mechanism of router cannot meet the requirements of now high performance network congestion control.In this dissertation we propose a congestion control architecture for high performance network, which improves the traditional congestion control architecture at TCP level and IP level, in order to resolve the problems of UDP strings and TCP-unfriendly strings possessing overabundance bandwidth which will easily result in network congestion ; at the same time, this improvements can resolve problems, such as global synchronization, lock out and full queue at a certain extent and improve the link utilization.FAST TCP changes the congestion measure mechanism of traditional TCP essentially and the detect mode of network capacity,and at the same time it improves the AIMD mechanism greatly. This dissertation researches the FAST TCP carefully, discusses its design motivation, architecture and the algorithm realization. We have embedded the FAST TCP protocol into Linux kernel 2.4.22 and designed some experiments in Giga Ethernet. Then we use Iperf tool to test the performance of FAST TCP and traditional TCP. The result has shown that the performace of FAST TCP is higher than traditional TCP.The SRED algorithm's early drop probability function curve changes suddenly which will result in the instability of queue length in router and dropping packets unnecessarily. We introduce a router queue law to analyse the limitation of SRED and design a new algorithm:E-SRED. We implement the E-SRED algorithm on NS-2.26 and design 4 experiments. The results validate that E-SRED algorithm can keep queue length more stable and make the queue length converge more quickly than SRED, especially under the heavy load.This dissertation has studied the architecture and flow control mechanism of IBM's network processor-NP4GS3 and proposed a project of how to implement the E-SRED algorithm on NP4GS3. We have designed two hardware functions of the algorithm based on the hardware assistant of NP4GS3 and discussed the parmaters of the hardware functions. At the same time, we have also presented the pseudo code of the project and analysed the feasibility of the project.