Abstract: The least glamorous problem in networking is the generation of synthetic traffic for network simulations. The implicit goal of synthetic traffic generation is to produce "realistic" network traffic, however, the definition of "realistic" is poorly understood, as are the factors that influence our ability to generate synthetic traffic that is representative of traffic observed on an actual link.
In order to perform valid experiments, network simulators require source-level synthetic traffic generators that correspond to valid, contemporary models of application and user behavior. Unfortunately, such traffic generators do not exist today as the Internet is evolving far more rapidly than our present ability to understand and model the mix and use of applications that account for the majority of bytes transferred on the Internet.
This talk presents a measurement-based method of synthetic traffic generation wherein a source-level model of the mix of TCP applications found on a network link is automatically generated without any knowledge of which applications are actually using the network. In addition, the network-dependent parameters of the experimental environment can be tuned to adjust the degree of realism in the synthetically generated traffic and, ultimately, to approach a high-fidelity reproduction of traffic observed on a network link. A case study is presented of reproducing in a laboratory testbed, important features of traffic found on two Internet links -- an OC-48 link in the Internet2 Abilene backbone and a 1 Gbps Ethernet link connecting the UNC campus with its ISP.
Combined, these results demonstrate methods and tools for increasing the level of realism in network simulations as well as understanding the structure and mix of traffic observed on network links.