Switches under real Internet traffic

We propose a novel methodology to generate realistic traffic traces to be used for the performance evaluation of switches. Indeed, real Internet traffic shows long and short range dependency characteristics, difficult to capture by flexible, yet simple, synthetic models. One option is to use real traffic traces, which, however, are difficult to obtain, as it requires traffic to be captured in different places, with synchronization and management problems. We therefore present a methodology to generate several synthetic traffic traces from a single real trace of packets, by carefully grouping packets belonging to the same flow in order to guarantee keeping the statistical properties of the original trace. After formalizing the problem, we solve it and apply the results to assess the performance of scheduling algorithms in high performance switches, comparing the results to other simpler traffic models traditionally adopted in the switching community. Our results show that realistic traffic degrades the performance of the switch by more than one order of magnitude with respect to traditional traffic models.