This paper analyzes how self-similarity is created in bottleneck buffers of computer networks. We argue that, in absence of heavy tailed section length or file size distributions, the sources which create self similarity in the network are certain buffers which are either too short or shared by too many TCP sessions. We analyse how congestion and long-range dependent traffic is generated in such a buffer shared by parallel TCP flows. We find that a single parameter, the ratio, determines the nature of the traffic leaving the buffer. If this ratio is above ≈ 3.0, TCPs stay in the congestion avoidance phase, if it is below then the congestion window dynamics becomes chaotic or stochastic. We show that in this phase the main properties of the traffic are determined by the exponential backoff mechanism. Emergence of chaotic TCP dynamics and long range dependence seem to be intimately related.