TCP/IP traffic dynamics and network performance: a lesson in workload modeling, flow control, and trace-driven simulations

The main objective of this paper is to demonstrate in the context of a simple TCP/IP-based network that depending on the underlying assumptions about the inherent nature of the dynamics of network traffic, very different conclusions can be derived for a number of well-studied and apparently well-understood problems in the area of performance evaluation. For example, a traffic workload model can either completely ignore the empirically observed high variability at the TCP connection level (i.e., assume "infinite sources") or explicitly account for it with the help of heavy-tailed distributions for TCP connection sizes or durations. Based on detailed ns-2 simulation results, we illustrate that these two commonly-used traffic workload scenarios can give rise to fundamentally different buffer dynamics in IP routers. Using a second set of ns-2 simulation experiments, we also illustrate a qualitatively very different queueing behavior within IP routers depending on whether the traffic arriving at the router is assumed to be endogenous in nature (i.e., a result of the "closed loop" nature of the feedback-based congestion control algorithm of TCP) or exogenously determined (i.e., given by some conventional traffic model --- a fixed "open loop" description of the traffic as seen by the router).