Detecting the points of network congestion is an intriguing research problem, because this information can benefit both regular network users and Internet Service Providers. This is also a highly challenging problem, because the Internet is designed to provide only end-to-end services, and its internals are in principal invisible to end users. Current techniques used to detect bottleneck positions have problems such as high probing overhead and low measurement accuracy. In this paper, we propose using Recursive Packet Trains (RPT) to detect the network congestion position. RPT combines two types of probing packets -- measurement packets and load packets -- in a single probing packet train. The idea is to let load packets generate a packet queue on the router, and to use the measurement packets at the beginning and the end of the train to measure the packet train length. By detecting the changes in the packet train length, we can derive the congestion points of the network path. RPT has the advantages that it only needs single-end control and that it has relatively low overhead. In this paper, we present the algorithm and evaluate it using both testbed experiments and Internet experiments.

18 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu