We revisit the TCP-modified Engset model proposed by Heyman et al. in [1]. The model deals with the superposition of a limited number of TCP connections alternating between file transmission and silence in a web-like fashion. We consider homogeneous sources only. (a) We take into account the effects of slow start and limited receiver window as well as small average file sizes. (b) We propose an alternative way for calculating the average connection rate in the superposition. (c) From the model we propose a way for calculating the queuing behavior; i.e., the overflow probability. (d) From this last point, we propose a new link buffer sizing rule. Comparison with extensive simulations shows that the average rate and duration, as well as, link utilization are accurately predicted for exponentially distributed file sizes. For longer tail distributions, the model remains accurate provided the receiver window is adjusted appropriately. The accuracy increases with increasing load. As concerns the queuing behavior, the same observation applies. Finally, the revisited model cannot be used to predict losses larger than about 1%. The model overestimates loss rates above that threshold.
