Forwarding devices

Abstract

Most popular simulation and emulation tools use high-level models of forwarding behavior in switches and routers, and give little guidance on setting model parameters such as buffer sizes. Thus, a myriad of papers report results that are highly sensitive to the forwarding model or buffer size used. Incorrect conclusions are often drawn from these results about transport or application protocol performance, service provisioning, or vulnerability to attacks. In this article, we argue that measurement-based models for routers and other forwarding devices are necessary. We devise such a model and validate it with measurements from three types of Cisco routers and one Juniper router, under varying traffic conditions. The structure of our model is device-independent, but the model uses device specific parameters. The compactness of the parameters and simplicity of the model make it versatile for high-fidelity simulations that preserve simulation scalability. We construct a profiler to infer the parameters within a few hours. Our results indicate that our model approximates different types of routers significantly better than the default ns-2 simulator models. The results also indicate that queue characteristics vary dramatically among the devices we measure, and that backplane contention can be a factor.