IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 6, JUNE 2004 569 TCP Retransmission Timeout Algorithm Using Weighted Medians Liangping Ma, Student Member, IEEE, Gonzalo R. Arce, Fellow, IEEE, and Kenneth E. Barner, Senior Member, IEEE Abstract—This letter presents a new retransmission timeout (RTO) algorithm based on recursive weighted median (RWM) filters for the transmission control protocol (TCP). The RTO algorithm utilized in current TCP implementations is Jacobson’s algorithm [1], which is based on recursive linear filtering. While linear filters are adequate for estimation in Gaussian signal envi- ronments, the round trip time (RTT) signals filtered to determine the RTOs are often impulsive. Thus, Jacobson’s algorithm is not effective in many cases. The proposed algorithm employs RWM filters that yield improved performance when operating on RTT signals with heavy tailed statistics. Simulation results show that the proposed method yields significantly tighter RTT bounds than Jacobson’s method over Internet traffic with heavy tailed statistics. Index Terms—Recursive weighted median filters, retransmission timeout (RTO), round trip time (RTT), transmission control pro- tocol (TCP). I. INTRODUCTION T RANSMISSION control protocol (TCP) provides reliable logical communications for applications (e.g., FTP, telnet, and HTTP) running on different hosts on the Internet [2, p. 82]. It is critical for TCP to have an accurate retransmission timeout (RTO) algorithm, which greatly affects data transfer delay and congestion control [3, p. 467]. An RTO algorithm assigns a timeout to each outgoing packet based on the past RTT obser- vations, where the RTT is the time it takes for a packet to travel from client to server and then back to the client [2, p. 88]. The desired RTO should be slightly larger than the corresponding RTT [1], [2, p. 224]. The RTO algorithm used in current TCP implementations was proposed by Jacobson [1]. More recently developed RTO algo- rithms include the Eifel Retransmission and the jitter-based al- gorithms [4], [5]. The Eifel algorithm was heuristically designed based on an assumed set of RTT characteristics, which may not hold for all network configurations. The jitter-based algorithm utilizes the fractal-like structure of the RTT processes, but a table has to be pre-configured for each Internet path making it unsuitable for real time applications. This letter proposes a new RTO algorithm based on recursive weighted median (RWM) filtering. The RWM algorithm yields better performance than Jacobson’s algorithm and is simple Manuscript received July 10, 2003; revised October 10, 2003. This work was supported in part by the National Science Foundation ITR-ANI under Grant 0312851. The associate editor coordinating the review of this manuscript and approving it for publication was Prof. Dimitris A. Pados. The authors are with the Department of Electrical and Computer Engineering, University of Delaware, Newark, DE 19716 USA (e-mail: lma@ece.udel.edu; arce@ece.udel.edu; barner@ece.udel.edu). Digital Object Identifier 10.1109/LSP.2004.827957 Fig. 1. Examples of RTT traces. Trace 1 (a) is for the connection pogo.udel.edu—clock.nc.fukuoka.jp, and trace 2 (d) is for the connection pogo.udel.edu—tictoc.tip.csir.au, both measured on April 18, 2001. A probe packet was sent every 60 s. (b) Zoom-in of (a), and (c) zoom-in of (b). Sample autocorrelation coefficients for both traces are depicted in (e), and the optimized weights for the RWM filter for trace 2 are illustrated in (f). enough to allow real time applications. The RWM approach is motivated by the following facts. First, Jacobson’s algorithm is not optimal. This algorithm is based on linear filtering, which provides adequate performance when the underlying signals exhibit Gaussian statistics. However, the RTT statistics are often impulsive [6], and cannot be accurately modeled as a Gaussian process, thus leading to poor RTO performance of Jacobson’s algorithm. The impulsive characteristic of the RTT is clearly shown in Fig. 1. Second, weighted median (WM) fil- ters are optimal for signals obeying the heavy tailed Laplacian distribution [7], [8]. Accordingly, WM filters yield improved performance when filtering signals with heavy tailed distribu- tions. A rich theory is available for the design and optimization of WM filters [7], [8]. Utilizing the robust characteristics of the WM filtering, the proposed RWM algorithm yields RTOs that are less susceptible to rapid variations on the data. The remainder of this letter is organized as follows. In Sec- tion II, Jacobson’s algorithm is discussed and the RWM algo- rithm is proposed in Section III. Simulation results and some comments are given in Section IV. II. JACOBSON’S RTO ALGORITHM In Jabobson’s algorithm, the RTT estimate is obtained by lin- early filtering the previous RTTs and the RTO is then set as the 1070-9908/04$20.00 © 2004 IEEE