370 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 11, NO. 3, JUNE 2003 Comparative Study of Various TCP Versions Over a Wireless Link With Correlated Losses Farooq Anjum, Member, IEEE, and Leandros Tassiulas Abstract—In this paper, we investigate the behavior of the various Transmission Control Protocol (TCP) algorithms over wireless links with correlated packet losses. For such a scenario, we show that the performance of NewReno is worse than the performance of Tahoe in many situations and even OldTahoe in a few situations because of the inefficient fast recovery method of NewReno. We also show that random loss leads to significant throughput deterioration when either the product of the square of the bandwidth-delay ratio and the loss probability when in the good state exceeds one, or the product of the bandwidth-delay ratio and the packet success probability when in the bad state is less than two. The performance of Sack is always seen to be the best and the most robust, thereby arguing for the implementation of TCP Sack over the wireless channel. We also show that under certain conditions the performance depends not only on the bandwidth-delay product but also on the nature of timeout, coarse or fine. We have also investigated the effects of reducing the fast retransmit threshold. Index Terms—Correlated losses, packet train model, perfor- mance analysis, TCP algorithm, TCP over wireless. I. INTRODUCTION T RANSMISSION Control Protocol (TCP) is the transport protocol used by many internet-based applications, including http, ftp, telnet, etc. TCP is a reliable end-to-end window-based transport protocol designed for the wireline networks characterized by negligible random packet losses. The way that TCP works is that it keeps increasing the sending rate of packets as long as no packets are lost. When packet losses occur, e.g., due to the network becoming congested, TCP decreases the sending rate. Thus, TCP infers that every packet loss is due to congestion and, hence, backs off in the form of reducing the send window. Extending TCP as used over the wireline links to the wireless links may not be an efficient solution due to the different characteristics of the wireline and the wireless links. This is because wireless networks are characterized by bursty and high channel error rates, unlike the wireline networks. Due to this, the throughput of a TCP connection over a wireless link suffers. In spite of this, the TCP protocol is still used to transfer data over the wireless Manuscript received June 3, 1999; revised April 3, 2002; approved by IEEE/ACM TRANSACTIONS ON NETWORKING Editor T. V. Lakshman. This work was supported in part by the National Science Foundation under CAREER Award NCR-9502614 and by the Air Force Office of Scientific Research under Grant 95-1-0061. This paper was presented in part at the ACM SIGMETRICS, Atlanta, GA, 1999. F. Anjum is with Telcordia Technologies, Morristown, NJ 07960 USA (e-mail: fanjum@telcordia.com). L. Tassiulas is with the Department of Electrical Engineering, Uni- versity of Maryland, College Park, MD 20742-3285 USA (e-mail: lean- dros@glue.umd.edu). Digital Object Identifier 10.1109/TNET.2003.813033 link, though a lot of attention is currently being given to the design of a better protocol over wireless links [2]–[4], [7], [15], [19]. Because of the difficulty of modeling the TCP protocol analytically, many of these studies have been simulation based. On the other hand, it is not possible to obtain insight into the effects of particular parameters on the behavior of TCP using simulations of specific settings. Further, investigations to improve TCP or design a better transport protocol can become less cumbersome given a simple and accurate analytical model for TCP. The first step toward the design of a better transport pro- tocol for wireless networks has to be a better understanding of the way TCP works over wireless links. This would reveal the reasons for the inefficiency of TCP over wireless links. There have been several efforts recently on the analytical study of TCP over wireline [12] as well as over wireless links [9]–[11], [20]. Two classes of random losses have been considered, indepen- dent identically distributed (i.i.d.) and correlated. The effect of i.i.d. packet losses on TCP performance is studied in [9]. They address the TCP versions Tahoe, Reno, and NewReno, but only in the context of a local network scenario. They also evaluate the various protocol features such as fast retransmit and fast re- covery. Unlike in this paper, the packet transmission times are assumed to be exponentially distributed, as is the transmission time of the packet on the lossy link. Further, they also model the congestion avoidance phase probabilistically in which each acknowledgment (ack) causes the window to be incremented by one with a certain probability. Note that both these assumptions have to be resorted to in this approach so as to carry out the mean cycle time analysis. In this study, the authors also consider the less frequent case where the size of the receiver window is the constraint on the sender’s window increase and not the bandwidth delay product of the link. Thus, this study precludes the study of the basic TCP mechanism whereby the window size is increased until there is a loss due to congestion. This loss is caused on account of the bandwidth delay constraint. In [13], Mishra et al. consider only OldTahoe over a link with i.i.d. losses. Lakshman and Madhow [11] consider Tahoe and Reno in a regime where the bandwidth-delay product of the network is high compared to the buffering in the network. They show using approximate analysis that random independent packet loss leads to significant throughput deterioration when the product of the loss probability and the square of the bandwidth-delay product is larger than one. In [10], Kumar and Holtzman consider the behavior of TCP Tahoe and OldTahoe in the presence of correlated packet losses. The results obtained are applicable only in case of very low bandwidth wireless links. More emphasis is placed on analyzing a link layer solution, of hiding the losses from the transport layer 1063-6692/03$17.00 © 2003 IEEE