The PFTK-model revised
Introduction
Transmission control protocol (TCP) is the de facto standard protocol for the reliable data delivery in the Internet. Recent measurements show that from 60% to 90% of today’s Internet traffic is carried by TCP [1]. Due to this fact, TCP performance modeling has received a lot of attention in the literature, see references in [2].
One of the most known and wide referenced analytical models of TCP throughput of a bulk transfer is the model proposed by Padhye et al. in [3], also known as the PFTK-model. This model describes steady-state throughput of a long-lived TCP Reno bulk transfer as a function of loss event rate, average round trip time (RTT), average retransmission timeout value (RTO), and receiver window size. It assumes a correlated (bursty) loss model that is better suited for FIFO Drop Tail queues currently prevalent in the Internet.
Unfortunately, this model does not capture slow start phase after timeout and uses simplified representation of fast retransmit/fast recovery dynamics in the presence of correlated losses as having negligible effect on TCP Reno throughput. As it will be shown later, such simplifications may lead to overestimation of TCP Reno throughput. Since new analytical TCP throughput models are often compared with the PFTK-model (e.g., [4], [5], [6]) and use its resultant formula (e.g., [7], [8]), such inaccuracy in throughput estimation can lead to inaccurate results or incorrect conclusions.
In this paper, we propose a simple and more accurate steady-state TCP Reno throughput prediction model. This is achieved by careful examination of fast retransmit/fast recovery dynamics in the presence of correlated losses and taking into consideration slow start phase after timeout.
The rest of the paper is organized as follows. Section 2 describes assumptions we made while constructing our model. Section 3 presents a detailed analysis of the proposed model. Section 4 describes model validation experiments, presents an analysis of the accuracy of our model and the one proposed in [3]. Finally, Section 5 concludes the paper.
Section snippets
Assumptions
The refined model we develop in this paper has exactly the same assumptions about endpoints and network as the PFTK model presented in [3]. We assume that the sender uses TCP Reno congestion control algorithm based on [9] and always has data to send. Since we are focusing on TCP performance, we do not consider sender or receiver delays and limitations due to scheduling or buffering. Therefore, we assume that the sender sends full-sized segments whenever the congestion window (cwnd) allows,
Model building
According to [9], segment loss can be detected in one of two ways: either by the reception at the sender of “triple-duplicate” ACK or via retransmission timeout expiration. Similarly to [3], let us denote the first event as a TD (triple-duplicate) loss indication, and the second as a TO (timeout) loss indication. As in [3], we develop our model in three steps: when the first loss indication in a cycle is exclusively TD (Section 3.1); when the first loss indication in a cycle is either TD or TO
Model validation through simulation
In order to validate the proposed model and compare it with the one presented in [3], we compared the results obtained from the both analytical models against simulation results obtained from ns-2 [13]. We performed experiments using the well-known single bottleneck (“dumbbell”) network topology. In this topology all access links have a propagation delay of 1 ms and a bandwidth of 10 Mbps. The bottleneck link is configured as a Drop Tail link and has a propagation delay of 8 ms, bandwidth of 2 Mbps
Conclusion and future work
In this paper we developed an analytical model for predicting TCP Reno throughput in the presence of correlated losses. The model is based on the one proposed in [3] and improves it by taking into consideration fast retransmit/fast recovery dynamics and slow start phase after timeout. The presented model has the average error smaller than 5% over a wide range of loss event rate values with the mean of 3%, while the one, proposed in [3] performs well when the loss event rate is quite small and
Roman Dunaytsev received his M.Sc. degree in computer science and telecommunication systems and Candidate of Science degree in telecommunication networks from St.-Petersburg State University of Telecommunications, Russia, in 1999 and 2005 correspondingly. Currently he is a researcher in the Institute of Communications Engineering at Tampere University of Technology, Finland, working towards his Ph.D. degree. His research interests include TCP performance modeling and real-time multimedia
References (16)
- et al.
An integrated model for the latency and steady state throughput of TCP connections
Performance Evaluation
(2001) - M. Fomenkov, K. Keys, D. Moore, k. claffy, Longitudinal study of Internet traffic in 1998–2003, in: Proceedings of...
- J. Olsen, Stochastic modeling and simulation of the TCP protocol, Ph.D. thesis, Uppsala University, Sweden, October...
- et al.
Modeling TCP Reno performance: a simple model and its empirical validation
IEEE/ACM Transactions on Networking
(2000) - S. Fortin, B. Sericola, A Markovian model for the stationary behavior of TCP, Technical Report RR-4240, IRISA-INRIA,...
- O. Bogoiavlenskaia, M. Kojo, M. Mutka, T. Alanko, Analytical Markovian model of TCP congestion avoidance algorithm...
- N. Cardwell, S. Savage, T. Anderson, Modeling TCP latency, in: Proceedings of IEEE INFOCOM, 1742–1751, March...
- S. Fu, M. Atiquzzman, Modeling TCP Reno with spurious timeouts in wireless mobile environment, in: Proceedings of 12-th...
Cited by (11)
A study of TCP performance in wireless environment using fixed-point approximation
2012, Computer NetworksCitation Excerpt :According to the structure of FPA-based models in order to come up with conclusions about accuracy of the model we need to investigate accuracy of both parts (see discussion in [1], Sections 4.2.5 and 4.2.6). Note that accuracy of Mathis and PFTK models has been checked in many studies (see e.g. [31,13] among many others). These models have been found to provide fairly accurate results for those loss environments they are intended for.
ECon: Modeling the network to improve application performance
2019, Proceedings of the ACM SIGCOMM Internet Measurement Conference, IMCDevelopment of a method for the experimental estimation of multimedia data flow rate in a computer network
2018, Eastern-European Journal of Enterprise TechnologiesModeling Multi-Path TCP throughput with coupled congestion control and flow control
2015, MSWiM 2015 - Proceedings of the 18th ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile SystemsConcurrent multipath transfer using SCTP: Modelling and congestion window management
2014, IEEE Transactions on Mobile Computing
Roman Dunaytsev received his M.Sc. degree in computer science and telecommunication systems and Candidate of Science degree in telecommunication networks from St.-Petersburg State University of Telecommunications, Russia, in 1999 and 2005 correspondingly. Currently he is a researcher in the Institute of Communications Engineering at Tampere University of Technology, Finland, working towards his Ph.D. degree. His research interests include TCP performance modeling and real-time multimedia traffic delivery and management.
Yevgeni Koucheryavy is a Senior Research Scientist in the Institute of Communications Engineering at Tampere University of Technology, Finland. Between 1995 and 2000 he was with St.-Petersburg Telecommunications R&D Institute holding various technical and managerial positions. Yevgeni actively participates in IST projects, in particular he chairs COST 290 “Wi-QoST: Traffic and QoS Management in Wireless Multimedia Networks”. Yevgeni has authored or co-authored over 50 papers in the field of advanced wired and wireless networking and teletraffic theory. He co-edited several proceedings books published in LNCS, Springer. Yevgeni serves on TPC of several conferences and workshops; he is a co-chair of WWIC 2005 and 2006, and TPC chair of NEW2AN 2004 and 2006. He also currently serves as an Editorial Board member for Elsevier Ad Hoc Networks Journal, IEEE Journal of Communication Software and Systems (JCOMSS). His current research interests include short-range wireless communications, real-time services traffic optimization and cross-layer techniques.
Jarmo Harju received his M.Sc. from Helsinki University of Technology in 1979 and Ph.D. in mathematics from the University of Helsinki in 1984. During 1985–89 he was a senior researcher at the Telecommunications Laboratory of the Technical Research Center of Finland, working with the development of protocol software. In 1989–95 he was professor of data communications at Lappeenranta University of Technology. Since 1996 he has been professor of telecommunications in the Institute of Communications Engineering at Tampere University of Technology, Finland, where he is leading a research group concentrating on network architectures and QoS issues.