Performance Evaluation of TCP Variants Over High Speed Satellite Links
Hiroyasu Obata, Kenji Ishida · 25th AIAA International Communications Satellite Systems Conference (organized by APSCC) · 2007
Bandwidth of wireless links becomes larger by the progress of wireless communication technology in recent years. Especially, the link speed of satellite networks will be improved like optical networks. For example, WINDS (Wideband InterNetworking engineering test and Demonstration Satellite) will be developed by JAXA (Japan Aerospace Exploration Agency) and NICT (National Institute of Information and Communications Technology) in Japan as a gigabit wireless satellite network. WINDS is a geostationary satellite and can communicate interactively at a rate of up to 1.2Gbps. However, there has been no evaluation on TCP performance by using such high speed satellite links. It is known that the most standard implementation of TCP congestion control method such as TCP-Reno performs poorly in satellite Internet due to its high bit error rate and long propagation delay. Therefore, we have been proposed a TCP congestion control method TCP-STAR for satellite communication systems. TCP-STAR is an end-to-end congestion control method and has three new mechanisms, namely Congestion Window Setting (CWS) based on available bandwidth, Lift Window Control (LWC), and Acknowledgment Error Notification (AEN). In this paper, we evaluate the throughput performance of TCP-STAR compared with the other TCP variants over a high speed satellite communication system such as WIDNS by a network simulator ns2. In simulation experiments, we compared TCP-STAR with TCP-WestwoodBR which is a typical variant for wireless and satellite networks, and TCP variants (FAST TCP, HighSpeed TCP, and Scalable TCP) for high speed optical networks. Simulation experiments show that TCP-STAR can obtain the best throughput comparing with other TCP variants when the bandwidth is 1.2Gbps. Moreover, we evaluate the throughput when bandwidth of data path and ack path is asymmetric. As a result, we found that sucient throughput corresponding to the data path bandwidth cannot be