Measuring and modeling the performance ofhigh-speed data transport protocols

Promotie: mw. J.S. Sansa-Otim, 11.00 uur, Academiegebouw, Broerstraat 5, Groningen

Proefschrift: Measuring and modeling the performance ofhigh-speed data transport protocols

Promotor(s): prof.dr. J.M. van der Hulst

Faculteit: Wiskunde en Natuurwetenschappen

Contact: Julianne Sansa-Otim, tel. 050-3634073, e-mail: sansa@astro.rug.nl

Measuring and modeling the performance ofhigh-speed data transport protocols

The widely used Transmission Control Protocol (TCP) for information transfer from computer to computer does not perform optimally in the case of high-speed networks. It either does not effectively use the available bandwidth or causes unnecessary congestion because of its simple control capabilities. The challenge in the era of increasing high-speed bandwidth capabilities and increased use of these capabilities is to improve the data transport protocols such that the available bandwidth is better utilised. Motivated by the requirements of real time Very Long Baseline Interferometry (VLBI) to exploit the standard available network we investigated the behaviour of different high-speed data transport protocols and embarked on network simulations to help evaluate existing and new protocols. Chapter 1 provides a brief description of a number of high-speed applications, reviews the coverage, capacity and quality of service offered by current high-speed networks, and discusses the future growth prospects of these networks. In addition we discuss previous efforts in enhancing TCP, outline the performance metric parameters used to evaluate them and the performance evaluation methodologies employed in this thesis to study the different protocols.

High-speed TCP protocols are not necessarily efficient and can be improved. Before considering improvements to existing protocols it is necessary to first examine their performance in real networks. Chapter 2 describes such experiments for the case of e-VLBI. The goal is twofold: (i) to identify the bottlenecks and analyse what aspect of the TCP protocols causes them; and (ii) to review a practical alternative to real experiments, i.e. network simulations, and examine which methods are appropriate for high-speed bandwidth modeling.

In Chapter 3 the performance of eight high-speed TCP protocols has been examined in a network simulation environment. One of these is a newly designed protocol (HTCP-BE) which combines the positive aspects of two existing protocols (H-TCP and TCPW). The goal was to quantify performance in terms of throughput and stability and identify the optimal protocol. It is shown that indeed some improvement can be made by combining the positive aspects of two protocols in a new one, HTCP-BE, as proposed in this chapter.

Discrete event simulators are very useful for studying network traffic behaviour, but become too computer intensive in the case of multiple flows in complex network configurations. An elegant solution is to describe network traffic as a fluid, using differential equations. In Chapter 4 we derive fluid flow models for three high-speed protocols and investigate whether these models adequately describe the behaviour of the high-speed protocols in the case of both low-speed and high-speed connections. We implemented both the Random Early Detection (RED) and Droptail (DT) queuing mechanisms in the fluid flow models and conclude that both can be well described. We then use the designed fluid flow models to compare the performance of six TCP protocols on high-speed links and examine throughput, responsiveness, stability and fairness. Finally we briefly discuss the limitations of the fluid flow models we designed and validated.

High-speed TCP protocols with delay-based congestion detection have received far less attention in comparison to their loss-based counterparts . The reason is that measuring delay is considered difficult and unreliable. Loss-based protocols on the other hand are intrinsically inefficient because they require losses to occur in order to adjust to the available bandwidth and therefore introduce irrecoverable loss of bandwidth for themselves and for all other flows. It would be much more elegant if protocols were able to reliably probe the available bandwidth by measuring the delays and adjust the throughput based on the delay experienced to avoid losses altogether. In Chapter 5 we therefore revisit the issue of measuring delay reliably and propose a refined method of measuring available bandwidth which we implement in a new TCP, which we named F-TCP. We then prove using network simulations that this new delay-based TCP is able to operate efficiently in high-speed networks and co-exists fairly, not only with flows using the same protocol, but also with flows using loss-based protocols. We show that F-TCP has good throughput efficiency, responsiveness, intra-protocol fairness and TCP friendliness properties. We also prove fair co-existence between a delay-based protocol (F-TCP) and a loss-based protocol (regular TCP), so F-TCP maintains its fair share of the link.

Finally Chapter 6 provides a summary and future outlook for the research presented in this thesis.