VFAST TCP: A delay-based enhanced version of FAST TCP Salem Belhaj, and Moncef Tagina Abstract—This paper is aimed at describing a delay-based end- to-end (e2e) congestion control algorithm, called Very FAST TCP (VFAST), which is an enhanced version of FAST TCP. The main idea behind this enhancement is to smoothly estimate the Round-Trip Time (RTT) based on a nonlinear filter, which eliminates throughput and queue oscillation when RTT fluctuates. In this context, an eval- uation of the suggested scheme through simulation is introduced, by comparing our VFAST prototype with FAST in terms of throughput, queue behavior, fairness, stability, RTT and adaptivity to changes in network. The achieved simulation results indicate that the suggested protocol offer better performance than FAST TCP in terms of RTT estimation and throughput. Keywords—Fast tcp, RTT, delay estimation, delay-based conges- tion control, high speed TCP, large bandwidth delay product. I. I NTRODUCTION Congestion appears in intermediate network nodes (eg. routers) when flow in is higher than flow out, i.e. the load is temporarily higher than what the resources are able to treat, which causes delay increasing, throughput reduction, and packet loss (which may be retransmit in the case of TCP, whereas the retransmission will not happen for UDP). In the ideal case, the congestion control algorithm aims at achieving TCP throughput is equal to the link bandwidth by setting a correct cwnd that reflects the available bandwidth, but this needs zero loss and zero congestion. Since the throughput is obtained by dividing the window by RTT, the congestion window is obtained as the product bandwidth-delay. Traffic sources dynamically adapt their sending rates in response to congestion measure in their paths. In reality, sources set their congestion windows dynamically and consequently their sending rates. On the Internet, this is performed by TCP Reno, and its variants, in source and destination hosts involved in data transfers. However, it is well known that the existing loss- based Additive Increase Multiplicative Decrease mechanism (AIMD) of TCP [1], [2] have been a critical factor in the high utilization of the available bandwidth, as the network infrastructure scales up in capacity. Moreover, recent real time applications, such as teleoperation and telemedicine, require access to high bandwidth real time data, with predictable or constant low-latency. Yet, an efficient and fair sharing of the network resources among competing flow users is desired. These propositions lead to several possible approaches for controlling best-effort traffic at large windows. The current S. Belhaj is with the Analysis and Control Systems Research Laboratory, National Engineering School of Tunis, ENIT, B.P. 37, 1002 Tunis, Tunisia e-mail: (Salem.Belhaj@ensi.rnu.tn). M. Tagina is with National School of Computer Sciences, ENSI, University of Manouba, 2010 la Manouba, Tunisia e-mail: (Moncef.Tagina@ensi.rnu.tn). Manuscript received November 14, 2008; revised December 11, 2008. TCP implementation underutilzes the available bandwidth over high-speed long distance networks [3], which has mo- tivated several recent proposals for congestion control of high bandwidth-delay networks. In this context, a lot of promising algorithms were proposed as the possible replacement of the current TCP at large windows, including HSTCP (HighSpeed TCP [3]), STCP (Scalable TCP [4]), FAST TCP (Fast AQM Scalable TCP [5]), BIC TCP [6], CUBIC [7], CTCP (Com- pound TCP [8]) and H-TCP [9]. The main issues which must be considered when developing such an e2e congestion control algorithm are, as listed in [10]: How often to change the congestion window (frequency)? and how much should the change be? The work presented in this paper is an extended version of our preliminary results concerning this improved version of FAST TCP [11]. The suggested VFAST protocol offer better RTT estimation and throughput performance than FAST TCP. This paper is organized as follows: Section II gives a short overview of some theoretical background relative to e2e delay and congestion control. Section III briefly reviews our model and the basics of delay-based congestion control algorithms. FAST TCP limitations are addressed in section IV. Our VFAST design is presented in section V. Simulation results are discussed in section VI. Finally, conclusions and perspectives for future work are presented in section VII. II. BACKGROUND A. The Internet end-to-end delay Communication over a best effort packet-switching network, such as the Internet, is characterized by random losses and random delays. Its dynamic is actually time-variant, and de- pends on the Quality of Service (QoS) factors like bandwidth, delays, losses, etc. The e2e delay is the sum of delays experienced at each hop from the source to destination. The delay encountered at each intermediate node may be seen as a sum of two principal components: a constant component which includes the propagation delay and the transmission delay, and a variable component which includes the processing and queuing delay. This last component is the major source of uncertainty for e2e delay estimation, because it depends on the instantaneous traffic. Various studies attempted to characterize e2e dynamics of the Internet [12], [13], [14]. In literature, RTT is often used to study the Internet dynamics [15], [16], which requires measurement only at one end. Alternatively, the One-way Transmission Time (OTT) needs the collabo- ration between sender and receiver side to obtain accurate International Journal of Computer and Information Science and Engineering 2;2 © www.waset.org Spring 2008 59