Network link dimensioning: a measurement & modeling based approach

Network link dimensioning is commonly performed through rules of thumb, using only five minute traffic averages. More precise network link dimensioning, however, requires a thorough insight into the interrelationship between (i) the traffic offered (in terms of the average load, but also its fluctuations), (ii) the desired level of performance, and (iii) the required bandwidth capacity. In this thesis, we propose procedures and formulas that facilitate precise network link dimensioning. The formulas that we develop estimate the bandwidth capacity that is required to achieve a desired performance level, for a given, general traffic model. Our measurements have indicated that the Gaussian traffic model shows an excellent fit. Then, in order to apply the bandwidth capacity formula, estimates are needed of the average traffic rate and the fluctuations at small timescales. The former statistics are easy to obtain, whereas the latter are substantially harder. We introduce a novel approach to estimate the traffic fluctuations at small timescales, without requiring measurements at these small timescales, by coarse-grained polling of the buffer occupancy. We extensively validate the traffic models, the formulas for the required bandwidth, and the novel method to estimate traffic fluctuations, using hundreds of detailed measurements of real network traffic that are collected from five different locations in the Netherlands.