We propose a method for performance modeling of TCP/IP over ATM. The modeling is focused on user level behavior and demands. The basic components of our model are the arrivals of new TCP connections according to a Poisson process, and file sizes following heavy-tailed distributions. Using simulations we investigate the impacts of the behavior of such a source on the traffic at lower layers in the network. The benefits of considering the whole system in this way are several. Compared to commonly suggested models operating solely on the link level, a more complete and thorough view of the system is attained. The model also lends itself easily to studies of improvements and modifications of the involved protocols, as well as new ways of handling the traffic. The verification of our model demonstrates that it captures relevant features shown to be present in traffic measurements, such as high variability over long time-scales, self-similarity, long-range dependence, and buffering characteristics.