On each panel, the probability of exceeding a particular queue length for 500 blocks of intervals are graphed against the block HTTP connection rates. Each row of panels has the same queue length and each column has the same run type. The smooth curve on each panel is a loess fit with normal local linear fitting and a span of 0.75.
As described in the paper "On the Nonstationarity of Internet Traffic", we carry out 500 queueing simulation experiments, one experiment per block. Each experiment is a $2^3$ factorial: three factors each at 2 levels, and 8 runs consisting of all possible combinations of the levels of the three factors. In Run 1, the packets are fed into the queue according to the observed inter-arrivals and sizes. In each of the other 7 runs, the sizes and inter-arrivals of the block are altered according to the factors.
The first factor is the form of the inter-arrival marginal distribution; we use either the original observed values or the values of an exponential distribution.
The second factor is the order of the inter-arrival times in the block; we use either the original order or a randomized order.
The third factor is the packet size. We take the sizes to be either the original ones or to be constant and equal to the mean size of the packets in the block.
Here is the complete queueing result plots (compressed postscript, pdf) from the experiments used in the paper "On the Nonstationarity of Internet Traffic".