Kalman filter-Based Approximation of a Network Calculus TCP Controller Mingyu Chen, Manohar N. Murthi and Kamal Premaratne Department of Electrical and Computer Engineering University of Miami, 1251 Memorial Drive EB406, Coral Gables, FL 33146 USA mchen@umsis.miami.edu, {mmurthi, kamal}@miami.edu Abstract— Delay-based TCP congestion avoidance methods such TCP Vegas, Enhanced Vegas, and Fast TCP all suffer from throughput degradation as illustrated by their inability to fully utilize the link bandwidth in a bidirectional bottleneck link topology. In this paper, we propose a Kalman filter implemen- tation of a TCP congestion avoidance scheme that is derived by approximating the ideal window controller generated from an approach based upon network calculus (NC) notions. With an improved bandwidth estimation method, this Kalman Filter- based NC controller presents a marked improvement over all other delay-based TCP congestion avoidance schemes. In ns-2 simulations, it is able to fully utilize the link bandwidth in both directions with the router queue levels achieving stable set points with reduced queueing delay variation under highly dynamic conditions. Index Terms— Congestion control, TCP Vegas, Fast TCP, Network Calculus, Kalman filter I. I NTRODUCTION D ELAY-BASED TCP congestion avoidance mechanisms such as TCP Vegas [2] and Fast TCP [4] detect network congestion in the early stages and prevent periodic packet loss that occurs in TCP Reno [1]. It has been demonstrated that TCP Vegas and Fast TCP can achieve much higher throughput than TCP Reno, [2], [4]. However, current delay- based TCP congestion avoidance methods suffer from several shortcomings. For example, both TCP Vegas and Fast TCP suffer from serious throughput degradation in the forward path (data trans- fer direction) when the reverse path (acknowledgement packet direction) is congested or otherwise highly utilized [3], [5], [6]. In the case of a simple bidirectional bottleneck link, one of the links is not transmitting data at its link capacity, indicating that the corresponding queue has settled at a queueing delay close to zero. This is largely due to the mis-estimation of the available bandwidth by the source window control methods. Several works have proposed remedies to this well-known problem of reverse path congestion. For example, the En- hanced TCP Vegas [3] follows the same basic controller strategy as TCP Vegas, but utilizes the time-stamp option and considers relative delay to improve parameter estimation The work reported herein is supported by the US National Science Foundation (NSF) via Grants CNS-0519933, CCF-0347229 and IIS-0325260. in the window controller, thereby improving the throughput performance of TCP Vegas. However, the throughput will still degrade as the backward traffic load increases, [3]. Moreover, as we will show in this paper, the router queue levels on a bidirectional bottleneck link oscillate between zero and a large value, not settling to a set point. Therefore, both the forward and reverse paths of a bottleneck link do not feature maximum link throughput for Enhanced TCP Vegas. In [6], we considered delay-based congestion avoidance within a Network Calculus (NC) setting. In particular, we demonstrated that TCP Vegas, Enhanced TCP Vegas and Fast TCP can be viewed as different approaches to approximating an optimal NC window controller, with each TCP variant making different assumptions in terms of parameter estimation and control implementation strategy. Furthermore, in [6], we addressed the throughput degradation problems of TCP Vegas and Fast TCP by deriving another optimal NC controller, termed the F-model, with a different control objective. In particular, the F-model NC controller attempts to buffer a fixed number of packets solely in the forward path between the source and destination nodes. In ns-2 simulations [6], an initial approximation of the F- model NC controller achieves performance gains in terms of link throughput and source node throughput fairness over TCP Vegas, Enhanced TCP Vegas, and Fast TCP. Although providing better performance than these other delay-based schemes, the approximated F-model controller in [6] does not always fully utilize the link bandwidth in both directions of a bidirectional bottleneck link. In this paper, we take another approach to approximating the ideal F-model NC controller, which is actually non-causal. In particular, we utilize the Kalman filter to approximate the F-model NC controller. With an improved bandwidth estimation and implementation method, the resulting practical window controller, in ns-2 simulations, is shown to possess several desired performance characteristics: Both links in a bidirectional bottleneck link are fully utilized; link through- put in one direction does not degrade due to congestion in another direction; the router queue levels in both directions roughly settle to stable set points, thereby providing smaller queuing delay variance to the data transfer application; the window controllers quickly adapt to highly dynamic network 591 1-4244-0350-2/06/$20.00 ©2006 IEEE