Summary

Our basic traffic generation method, source-level trace replay, results in highly realistic synthetic traffic. This method is however inflexible, in the sense that the same connection vectors are started at the same relative times in every replay. In this chapter, we proposed two methods for resampling an original trace of connection vectors, to create a new trace with similar statistical characteristics. This similarity is defined in terms of source-level behavior and network-level parameters, so the resampling methods also modify connection vector start times. Our first resampling method is Poisson Resampling, which chooses connections vectors at random and assigns them exponentially distributed inter-arrival times. Our measurement results demonstrated that this choice of the inter-arrival distribution is appropriate, in the sense that the marginal distribution of the connection inter-arrival in every trace we examined is remarkably consistent with the exponential distribution. Our second resampling method is Block Resampling, which chooses blocks of connection vectors at random. Unlike Poisson Resampling, Block Resampling preserves the dependency structure of the original connection arrival process. This makes it possible to reproduce the moderate long-range dependence that we observe in the connection arrivals of our traces.

Besides presenting two resampling methods, we also studied how to control the offered load by a resampled trace. Firstly, we demonstrated that the number of connections and the average offered load are not strongly correlated. This means that controlling the number of connections in the resamplings does not provide a good way of creating resampled traces with a specific target offered load. This is a common requirement when a set of experiments covers a range of offered loads in an empirical study. In order to address this difficulty, we propose to drive the resampling by a target total size of the ADUs in the resampling rather than by a target number of connections. We used this approach to develop byte-driven versions of Poisson Resampling and Block Resampling, which are shown to result in highly predictable offered loads.