INDEX
| Raweb 2003 / Project-Team : reso
|
HELP INDEX |
|
Key words: High speed transport, congestion control, flow control.
The new congestion control solution we propose for high speed network is based on back-pressure flow control. Our Network of Queues proposal [66] suggests an outright departure from current TCP standards for some particular networks. The idea is to replace the current end-host-based TCP /IP congestion management by a network of flow-controlled links, according to a scheme known as ``back-pressure''. The challenge is to control the flow queue by queue with the functionalities already present in the intermediate networking equipment. The idea is to activate the 802.3x flow control in very high speed Ethernet links to reduce the feedback loop and to efficiently prevent the congestion. It has been proved theoretically that the back-pressure approach is better than classical end system feedback control approach. The issue to solve is to prove its validity in the actual equipments and in operational IP networks and to solve cross layer issues. Our proposal argues that, in some specific networking contexts like those of grids, using back-pressure as an addition to existing TCP /IP /Ethernet networking hardware and software may offer a valuable tradeoff between performance gain and migration cost. In order to develop insights on how the current network hardware and software behave relative to flow control, we forced a 100 Mb/s bottleneck in local gigabit testbed. The result of the first, basic experiment is a sawtooth-shaped throughput curve. When several hosts compete for the same bottleneck, the cooperative AIMD algorithm of TCP gives an approximately fair share of the capacity to each flow. The first promising conclusions are that this approach is feasible in a IP/802.3x environment and offers a smoother reaction to congestion compared to TCP or HS-TCP and a rapid convergence to fairness.
Key words: Performance measurement, Packet Pair methods, hop by hop capacity discovery.
To discover the characteristics of a path in terms of hop by hop capacity and utilization rate, we have proposed a new method and a tool. Our approach, using a hop-by-hop packet pair method and a fine analysis of the measurements, provides such information. The method consists in using the dispersion of a packet pair because it has many advantages compared with the Variable Packet Size method [79] . Cross-traffic taints the dispersion measurements with noise, which forces to elaborate complex analysis methodologies [60] . There are two kinds of errors : the first one is typically due to cross-traffic when packets are inserted between the two probes and hence the capacity is underestimated. The method we propose lies on an incremental discovery of the path characteristics. For this, we evaluate three parameters for each measurements distribution. The maximal mode is the easiest to determine. It corresponds to the interval with the maximum numbers of samples. The previous mode is the mode of the current distribution which has the same capacity value as the one estimated for the previous hop. The new mode is the mode with a capacity value strictly lower – a new mode implies that the capacity decreases – than the previous mode and which includes a sufficient number of samples (here 1% of the total number of measurements). For the first hop, the previous and new mode are the same. We evaluate the noise area too. This is defined as the little capacity values area which contains three or more side-by-side modes, i.e. not separated by an interval of at least a distribution step. This proposition has been validated in simulation, then implemented in Linux and validated experimentally. We have compared our method with others to define its limits and the potential utilizations on the developed tool. We have shown that this method is relatively non-intrusive, robust, relatively accurate and reliable and keep these qualities under bad network conditions (high load, long path, etc.) [32] [48] , [65] . We have shown that our tool works up to 1 Gbit/s [48] . We have validated the Linux implementation and have demonstrated that it provides usable results in real life, without the participation of the receiving computer and path routers. Results show that TraceRate can also estimate path utilization rate. We are actually studying a new data analysis method that can rapidly extract such aggregate information. We will explore how such precise information can be used by a transport protocol to better control transmission rate and end to end transfer delay. The fact that our tool can give in a single and non-intrusive measure the capacity and the available bandwidth is very promising.
|
|
|