Pathload: a measurement tool for end-to-end available bandwidth

Thu, 03/27/2008 - 14:27 by Damien Saucez • Categories:

The paper "Pathload: a measurement tool for end-to-end available bandwidth" presented at the PAM'02 conference by M. Jain and C. Dovrolis presents an tool to estimate the available end-to-end bandwidth.

The Pathload tools is build on the Self-Loading Periodic Streams (SLoPS) (see [08-03-17]).

After a short introduction to bandwidth theory, the authors presents PathLoad. The idea is simple. A sender periodically sends trains of timestamped packets. The receiver then determines if there is a increasing trend in the one-way-delay.

Basically, if no increasing trend in packets is detected by the receiver then the sending rate is less than the available bandwidth (i.e., the path has not been saturated by the flow). On the opposite, if an increasing trend is detected in the one-way-delay, the path is saturated and the sending rate is more important than the available bandwidth. On the first case, the rate is increased for the next experiment. On the second, the rate is reduced. This technique leads to a convergence around a region which is an approximation of the available bandwidth.

To estimate if the rate is more important than the avail-bw, the authors introduce two metrics. First, the Pairwise Comparison Test (PCT) gives the fraction of consecutive one-way-delay that are increasing. Second, the Pairwise Difference Test PDT) determines the start-to-end delay variation relative to the variation within the stream.

To determine if the delay increases, a tradeoff has been proposed on these metrics. If the result is higher than the tradeoff, the delay is marked as increasing for the current experiment.

To evaluate the tool, the authors measured the avail-bw with Pathload and compare the results with the MRTG data they had for the paths they measured. Based on the experiments, Pathload seems to be accurate.

Pathload sends the probes on UDP and use TCP as a control channel between the source and the destination. Pathload supposes that paths are not changing during the measurements.

Paper available at http://www.cc.gatech.edu/fac/Constantinos.Dovrolis/Papers/pam02.ps