Poisson versus Periodic Path Probing (or, Does PASTA Matter?)

The well-known PASTA (``Poisson Arrivals See Time Averages'') property states that, under very general conditions, the fraction of Poisson arrivals that observe an underlying process in a particular state is equal, asymptotically, to the fraction of time the process spends in that state. When applied to network inference, PASTA implies that a Poisson probing stream provides an unbiased estimate of the desired time average. Our objective is to examine the practical significance of the PASTA property in the context of realistic RTT, loss rate and packet pair dispersion measurements with a finite (but not small) number of samples. In particular, we first evaluate the differences between the point estimates (median RTT, loss rate, and median dispersion) that result from Poisson and Periodic probing. Our evaluation is based on a rich set of measurements between 23 PlanetLab hosts. The experimental results show that in almost all measurement sessions the differences between the Poisson and Periodic point estimates are insignificant. In the case of RTT and dispersion measurements, we also used a non-parametric goodness-of-fit test, based on the Kullback-Leibler distance, to evaluate the similarity of the distributions that result from Poisson and Periodic probing. The results show that in more than 90% of the measurements there is no statistically significant difference between the two distributions.