Network testbeds and simulators remain the dominant platforms for evaluating networking technologies today. Central to the problem of network emulation or simulation is the problem modeling and generating realistic, synthetic Internet traffic as the results of such experiments are valid to the extent that the traffic generated to drive these experiments accurately represents the traffic carried in real production networks. Modeling and generating realistic Internet traffic remains a complex and not well-understood problem in empirical networking research. When modeling production network traffic, researchers lack a clear understanding about which characteristics of the traffic must be modeled, and how these traffic characteristics affect the results of their experiments. In this dissertation, we developed and analyzed a spectrum of empirically-derived traffic models with varying degrees of realism. For TCP traffic, we examined several choices for modeling the internal structure of TCP connections (the pattern of request/response exchanges), and the round trip times of connections. Using measurements from two different production networks, we constructed nine different traffic models, each embodying different choices in the modeling space, and conducted extensive experiments to evaluate these choices on a 10Gbps laboratory testbed. As a result of this study, we demonstrate that the old adage of garbage-in-garbage-out applies to empirical networking research. We conclude that the structure of traffic driving an experiment significantly affects the results of the experiment. And we demonstrate this by showing the effects on four key network performance metrics: connection durations, response times, router queue lengths, and number of active connections in the network.