Statistically Accurate Network Measurements
Seminar Room 1, Newton Institute
Estimations of Internet loss rates from measurements often have large statistical errors. Any meaningful
application of Internet loss measurement therefore needs to incorporate these uncertainties
in the results. However, estimating measurement errors is difficult as it requires the knowledge of
the loss process itself. Many experiments use a simple and unrealistic Bernoulli model for the loss
process and severely underestimate these errors.
Here we develop SAIL a measurement tool that can accurately estimate not only the loss rate of
an end-to-end path but also its statistical errors. The key idea in our method is to capture the
correlation between packet losses using an ON/OFF renewal model and to use a Hidden Semi-
Markov Model algorithm to infer the parameters of the ON and OFF periods from measurement
data. Once the model parameters are known, statistical properties of the loss process such as the
loss rate, its variance, and the distribution of the loss and no-loss bursts can be easily computed.
This is co-work with Hung Nguyen.