| Abstract: |
Simulations with web traffic usually generate input by sampling a heavy-tailed object size distribution. As a consequence these simulations remain in transient state over all periods of time,
i.e. all statistics that depend on moments of this distribution, such as the average object size or the average user-perceived latency of downloads, do not converge within periods practically
feasible for simulations. We therefore investigate whether the 95-th, 98-th, and 99-th latency percentiles, which do not depend on the extreme tail of the latency distribution, are more suitable
statistics for the performance evaluation. We exploit that corresponding object size percentiles in samples from a heavy-tailed distribution converge to normal distributions during periods
feasible for simulations. Conducting a simulation study with ns-2, we find a similar convergence for network latency percentiles. We explain this finding with probability theory and propose a
method to reliably test for this convergence. |
