Project Description
People
Partners
Demonstrations
Press and Media
Grid Café

High Performance Networking

Data transport

The Transport Control Protocol (TCP), which is today the most common solution for reliable data transfer over the Internet, is not well adapted to the needs of Grid operations.

For example, TCP can penalize the overall performance of high performance flows to a much higher degree than the lower performance ones. In addition, in order to fully exploit high link capacity, the congestion avoidance algorithms of TCP, which depend on the congestion window size, have to be properly tuned. However, large window sizes produce bursts of output traffic, which, if the router queues are nosufficiently large to handle these bursts, may cause packet loss and a consequent poor user-perceived performance.

Although some ad-hoc approaches such as striped ftp allow to partly circumvent the problem of large window sizes, it is far from clear that these solutions will scale to Gigabit per second file transfer. It is also doubtful that existing TCP congestion avoidance approaches (i.e. Reno/Newreno) are well suited in a high-speed environment. Therefore, a number of different aspects need to be investigated such as:
· Performance of different TCP implementations,
· TCP auto-tuning mechanisms and alternative transport protocols,
· The relationship between QoS policies and throughput,
· Queue management techniques for congestion avoidance like Explicit Congestion Notification (ECN) and Random Early Detection (RED).

There are a number of ongoing projects, e.g. Web100, which are directly tackling the issue of poor file transfer performance over very high speed long distance networks, as well as better instrumentation of TCP/IP, in order to ease the tuning or even allow self-tuning of the applications. Activities will be carried out in collaboration with such projects, if possible, and will be based on relevant results achieved by the whole community, which includes among the others researchers from the University of Illinois Chicago (Electronic Visualization Lab), iCAIR (international Centre for Advanced Internet Research) & NorthWestern University.

End-to-end inter-domain QoS

The need for QoS has been recognised for some time, and various technological components can be used to implement this. For example the Differentiated Services architecture, one of the most recent approaches under definition at IETF, is based on different types of traffic treatment ranging from EF (expedited forwarding) to AF (assured forwarding) with different so called, code points. EU Projects already exist to demonstrate such QoS (e.g. SEQUIN), and commercial network providers already routinely provide such services within their administrative domain.

From the point of view of the Grid, however, end-to-end QoS integrated across a WAN comprising of multiple independent administrative domains has not yet been demonstrated routinely. At each domain boundary a common service level specification (SLS) has to be agreed upon, so that all the Grid domains can support a consistent end-to-end service despite the fact that each domain can choose a service implementation completely independent from the one adopted by other domains. Several research networks like GÉANT, Abilene, GARR, UKERNA and others are already in the process of defining or offering pilot services. Cooperation with the Internet Service Providers who provide testbed connectivity will be fundamental to speeding up the introduction of QoS features into production networks and to provide help with the definition of services and their validation. Thus the entire sector of moving QoS onto the heterogeneous WAN is highly innovative at this time. This project provides an ideal context in which to integrate these pilot QoS services into the middleware, and test them using different traffic classes.

Advance reservation

One of the fundamental Grid services consists of the possibility to reserve resources according to the application requirements. We call this advance reservation. Several issues related to this need further investigation. First of all, network resource reservation has to rely on mechanisms for Quality of Service support, for this reason, Grid domains may in principle adopt different mechanisms to address this need. The goal of this project is to ensure that seamless end-to-end advance reservation is possible. This is highly innovative. A second issue is related to the need to perform admission control so that only entitled applications can actually reserve network resources and routers have the possibility to reject illegal requests