This paper discusses the impact of path selection on end-to-end performance. In the internet, packets are sent from one host to another. The routing is a black box to clients. The routing path has latency, packet loss rate, and bandwidth. Which path is selected affects the routing performance. There is little research on how the performance is affected.
The key idea of this paper is to compare the default path selected by the routing protocol with several alternatives which are best paths selected based on different metrics. The best path is selected as following: first, paths between pairs of hosts are measured, such as round-trip time, loss rate, and bandwidth. Then a synthetic topology graph is generated based on the long-term averages. The calculated values are weights of edges in this graph. Finally, best alternative paths are selected through this graph. In this way, authors can have the difference between the default path and best alternates. Their experiments show that the default path is usually not the best in terms of latency, loss rate and bandwidth. The effect is more obvious during peak hours.
The Internet path selection is not perfect because at each hop the routing decision is based on local information. There is no a complete knowledge of the whole topology. The synthetic topology graph, on the other hand, provides sufficient information for path selection based on statistic data. It is possible to avoid congested queues at each host and select a path with less propagation delay.
The authors try to quantify the impact of path selection on performance. However, the quantification may not be accurate. The analysis is based on long-term averages, which may not accurately reflect the dynamic situation. Moreover, the datasets are collected from 1995~1999. Since Internet develops very quickly during the years, the complexity of topology, the quality of links and the traffic load vary greatly. More different network devices and OSs are connected to the Internet. The dataset collected in different years may not comparable to each other.
There are two open questions: is the conclusion/results apply to large populations? Since it is hard to change the protocols or policies, is there a chance to improve the path selection without violation routing policies?
Received on Wed Nov 02 2005 - 18:59:47 EST
This archive was generated by hypermail 2.2.0 : Wed Nov 02 2005 - 20:30:13 EST