In this paper we study the statistics of the realized throughput of elastic
document transfers, accounting for the way network bandwidth is shared dyn
amically between the randomly varying number of concurrent flows. We first
discuss the way TCP realizes statistical bandwidth sharing, illustrating es
sential properties by means of packet level simulations. Mathematical flow
level models based on the theory of stochastic networks are then proposed t
o explain the observed behavior. A notable result is that first order perfo
rmance (e.g., mean throughput) is insensitive with respect both to the flow
size distribution and the flow arrival process, as long as "sessions" arri
ve according to a Poisson process. Perceived performance is shown to depend
most significantly on whether demand at flow level is less than or greater
than available capacity. The models provide a key to understanding the eff
ectiveness of techniques for congestion management and service differentiat
ion.