Wireless Traffic: The Failure of CBR Modeling 

Stefan Karpinski, Elizabeth M. Belding, Kevin C. Almeroth
Department of Computer Science 
University of California
Santa Barbara, CA 93106-5110

Performance predictions from simulations in wireless networking 
rarely seem to match the behavior observed once the same 
technologies are deployed. We believe that one of the major 
factors hampering researchers  ability to make more reliable 
forecasts is the inability to generate realistic experimental 
workloads. To redress this problem, we take a fundamentally new 
approach to quantifying the realism of wireless traffic models. 
In this approach, the realism of a model is defined directly in 
terms of its ability to accurately reproduce the performance 
characteristics of actual network usage. This direct approach 
cuts through the Gordian knot of deciding which statistical 
features of traffic traces are significant. The first major 
contribution of this work is this new definition of workload 
realism, together with the analytical and statistical methodology 
to rigorously assess whether synthetic traffic models meet the 
definition. The second contribution is the conclusion that 
commonly used models of wireless traffic distort important 
metrics for performance evaluation at every layer of the wireless 
protocol stack. We show by example that this distortion can 
completely invert the relative performance of protocols. The last 
and most important contribution is the complete collection of 
ideas, techniques and analytical tools that will allow the 
development of more realistic synthetic traffic models in the 
future.