First International Workshop
on Protocols for Fast
Long-Distance Networks

February 3-4, 2003
CERN, Geneva, Switzerland









Antony Antony, NIKHEF, The Netherlands, Johan Blom, Cees de Laat, Jason Lee, and Wim Sjouw, University of Amsterdam, The Netherlands

Title: Characterization of TCP Flows over Large-Fat Networks


Most recent research and development in the area of high speed TCP has focused on the steady state behavior of TCP flows. However our experience with one of the first research only trans-Atlantic Lambda links (2.5 Gbit/s) and one of the first trans-Atlantic 10 Gbit/s Ethernet paths has clearly revealed the need to focus on the initial stages of TCP. The work we present here examines the behavior of large TCP flows over high bandwidth delay networks at the microscopic level. This inspection has led us to question whether the underlying network can truly support bursty protocols such as TCP at these very high speeds combined with such high latency. We will briefly describe the requirements needed to support high-speed single stream TCP in such an extreme network environment, without resorting to using multi-streams We will present experimental results from using these Trans-Atlantic links at iGrid2002 where we tuned various host parameters and used modified TCP stacks to try to improve performance over these links, and show how they differ with what the theoretical throughput is on these links.

Grid applications in general can be demanding in terms of bandwidth requirements. A typical large-scale scientific experiment involves at least two parties; a data producer and a data consumer. High-energy physics experiments such as LHC, D0 and EDG are good examples of this. The data for these projects is predominantly produced at just few locations, the data producer, and then several researchers at diverse geographical locations analyze this data, the data consumers. The large distances that often geographically separate the researchers makes it difficult to have the basic access to the data in a reasonable timeframe. For example, EDG roughly estimates its network requirement at an average sustained flow of 400 Mbps for a period of four months out of a year. Not only does this data need to be collected, but additionally a large part of it needs to be transferred between various locations, and this has to be done over the network. Network. architectures are evolving to meet this new and unprecedented demand.

We present here research into several new architectures and explore the scalability of protocols such as TCP to try to determine how applications can possibly attain the required high bandwidths needed. In the fall of 2001 SURFnet provided a 2.5 Gbit/s Lambda and a 10 Gbit/s Ethernet link between Amsterdam and Chicago for research along with several other research only links (DataTAG). Initial tests showed that increasing the speed of links and routers in the path alone was not sufficient to obtain an end-to-end high-speed data transfer. In fact the opposite was true, with this high bandwidth delay link, we were getting less the 10 Mbit/sec. We have since conducted extensive experiments on this Trans-Atlantic link both to understand how transport protocols behave and what additional requirements high speed flows, such as TCP, impose on these kinds of networks. One architectural shift in such high-speed networking is to minimize the number of routers (devices which process packets at layer 3 and above), and instead delegate packet forwarding to switching devices at layer 2 or below. Initial throughput measurements of a single TCP stream over such an extreme network infrastructure (i.e. the SURFnet Lambda) has shown surprisingly poor results. This has led us to study the dynamics of TCP at the microscopic level to better understand its behavior. The primary motivation of this work is the demand from HEP community to get maximum throughput over long distance links. Currently HEP cannot effectively use the available bandwidth and simply increasing network capacity does not improve end-to-end performance. The exclusive availability research only high-bandwidth delay networks and lambda’s has allowed us to approach problem this problem in a purely research oriented fashion.


PFLDnet 2003 is organized at CERN and sponsored by: