Performance Analysis of TCP with Nonpersistent Sessions

Most analytical models of TCP assume that the sessions have an infinite amount of data to transmit. In practice, it is well known that sessions arrive, transfer a finite amount of data and depart. This paper is about an analytical model for calculating the average bandwidth shares obtained by Additive Increase Multiplicative Decrease (AIMD) controlled ephemeral sessions that share a single (bottleneck) link. In particular, we consider the following traffic model: file transfer requests arrive in a Poisson process, each request requiring the transfer of an independent and exponentially distributed volume of data, which we view as a fluid. After transferring its required volume of data, each session departs. We develop an analysis of this model, and we identify a Markov regenerative structure in the rate evolution process. We provide a complete and exact analysis that yields the average bandwidth share obtained by sessions. We obtain numerical results from the analysis and from an ns simulation of TCP controlled sessions, and discuss the region of applicability of such a modelling and analysis approach. We find that the detailed behavior of TCP’s congestion avoidance phase is well characterised by our model. The analysis compares well with simulations for low offered loads and when the mean file transfer requests are large compared to the round-trip pipe. We observe that slow start has a considerable effect on performance of TCP controlled ephemeral sessions. We also find that our analytical model yields results that are close to those obtained from a processor sharing model with the link capacity taken to be 0.75 of its actual value. Keywords— TCP performance modelling, bandwidth sharing, processor sharing.