IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 18, NO. 7, JULY 2000 1289 Throughput Analysis of TCP on Channels with Memory Michele Zorzi, Senior Member, IEEE, A. Chockalingam, Senior Member, IEEE, and Ramesh R. Rao, Senior Member, IEEE Abstract—The focus of this paper is to analyze the relative sen- sitivity of the bulk throughput performance of different versions of TCP, viz., OldTahoe, Tahoe, Reno, and New Reno, to channel er- rors that are correlated. We investigate the performance of a single wireless TCP connection in a local environment by modeling the correlated packet loss/error process (e.g., as induced by a multi- path fading channel) as a first-order Markov chain. A major con- tribution of the paper is a unified analytical approach which allows the evaluation of the throughput performance of various versions of TCP. The main findings of this study are that 1) error corre- lations significantly affect the performance of TCP, and in partic- ular may result in considerably better performance for Tahoe and NewReno; and 2) over slowly fading channels which are charac- terized by significant channel memory, Tahoe performs as well as NewReno. This leads us to conclude that a clever design of the lower layers that preserve error correlations, naturally present on wire- less links because of the fading behavior, could be an attractive al- ternative to the development or the use of more complex versions of TCP. Index Terms—Bursty errors, Markov channel, slow fading, TCP. I. INTRODUCTION D UE TO rapid advances in the area of wireless commu- nications and the popularity of the Internet, provision of packet data services for applications like e-mail, web browsing, and mobile computing over wireless channels is gaining importance. Transport Control Protocol (TCP) is a reliable, end-to-end, transport protocol that is widely used to support applications like telnet, ftp, and http [1]. TCP was designed primarily for wireline networks where the channel error rates are very low and congestion is the primary cause of packet loss [2]. Since its original deployment, several modifications to TCP, including Reno, NewReno, and Vegas, have been proposed and their performance analyzed in wire- line networks [2]–[4]. Reno’s loss recovery algorithm is opti- mized for the case when a single packet is lost in a window of data. Hence, Reno can suffer performance problems when mul- tiple packets are lost in a window [3]. NewReno addresses this Manuscript received December 7, 1998; revised December 13, 1999. This work was supported in part by the Center for Wireless Communications, Uni- versity of California, San Diego. M. Zorzi is with the Dipartimento di Ingegneria, Università di Ferrara, 44100 Ferrara, Italy (e-mail: zorzi@ing.unife.it). A. Chockalingam is with the Department of Electrical Communication Engineering, Indian Institute of Science, Bangalore-560012, India (e-mail: achockal@ece.iisc.ernet.in). R. R. Rao is with the Department of Electrical and Computer Engineering, University of California, San Diego, La Jolla, CA 92093-0407 USA (e-mail: rrao@ucsd.edu). Publisher Item Identifier S 0733-8716(00)05287-2. problem by improving the loss recovery phase to handle this situation [3], [5]. However, the performance of these enhanced TCP versions on wireless fading links has not been adequately studied so far. There have been several recent investigations on different facets of wireless TCP [5]–[13]. Most of these studies do not consider the effect of correlation in multipath fading [14], [15]. Errors are often assumed to occur independently and with the same probability on each packet. Yet, because the multipath fading process in a mobile radio environment can be slowly varying for typical values of carrier frequency and user speed, the dependence between errors in the transmissions of consec- utive packets of data cannot be neglected. Although motivated by the behavior of the wireless channel, the loss model consid- ered here also applies to any environment where the packet loss process exhibits memory, e.g., due to congestion. Related work has been presented in [12], where the perfor- mance of the OldTahoe and Tahoe versions of TCP is analyzed assuming a two state Markov channel model. A simplified an- alytical model for the throughput of the Tahoe version in the presence of Markovian packet losses has been presented in [15], where it was shown that correlation has a beneficial effect. An open question in this regard is the relative difference between the performance of various versions of TCP over the fading channel. In this paper, we compare the performance of OldTahoe, Tahoe, Reno, and NewReno. In each instance, a single TCP connection is assumed to extend over a multipath fading channel modeled as a first-order Markov packet error model, as in [12] and [16]. We assume that a large data file is to be transferred from the base station to a mobile terminal, over a 1.5 Mbps wireless link that is characterized by very low delay-bandwidth product. As in [5], we assume instantaneous ACK’s. The assumption of a single TCP connection, with the TCP end points being at the base station and at the mobile terminal, is consistent with IS-99 [17], [18]. Two main findings of this paper are that 1) error correla- tions significantly affect the performance of TCP, and in par- ticular result in considerably better performance for Tahoe and NewReno; and 2) over slow fading channels, Tahoe performs as well as NewReno. This leads to the conclusion that a clever design of the lower layers that preserves error correlations, nat- urally present on wireless links because of the fading behavior, could be an attractive alternative to the development or the use of more complex versions of TCP. Another interesting observa- tion that can be drawn from the results presented in this paper is that using different versions of TCP, or even the same version but with different parameters, may lead to significant changes 0733–8716/00$10.00 © 2000 IEEE