Dimensioning Drop-tail and AQM (RED) buffers at access networks for optimal performance with bulk data TCP traffic

• Authors: Zlatka Avramova;Danny De Vleeschauwer;Sabine Wittevrongel;Herwig Bruneel
• Venue: Computer Communications
• Year: 2010

Dimensioning buffers is still an area open to research efforts and discussions and much study work has been conducted on this topic recently. Most of this work is focused on sizing buffers in core routers in which a huge number of flows compete. Different, sometimes contradictory theories and buffer dimensioning formulae are put forward, which take into account statistical properties when aggregating many TCP flows. In our study we also work under the widely accepted assumption that persistent TCP flows account for most of the buffer demand and short-lived (e.g. web-traffic) TCP or non-elastic (e.g. streaming) flows just add a little bit more to the buffer requirement. In the research dedicated to core routers, these are assumed as bottlenecks, which, given the high-bandwidth backbone networks today, is not very frequent. Very often, however, the narrow bandwidth last-mile network is the bottleneck for a TCP connection. Therefore, we explore buffers in the access network equipment (e.g. in a DSLAM access module), in which a few long-lived TCP-controlled connections compete, synchronization between flows is a frequent phenomenon, and we elaborate on the impact of buffer configuration on the access (down-) link performance. In this paper, a formula is presented for dimensioning access network (bottleneck) buffers for no loss on the flows competing through them, given the flows' parameters are known (number of flows, propagation delays, and advertized windows). The formula is validated against simulations. Further, we explore the sensitivity of this buffer dimensioning function to its parameters. As we show, in order to avoid buffer-incurred losses, the required buffer becomes easily very large, of the order of hundreds of packet units only for a few flows. This is not feasible for equipment implementation, nor is desirable because of other side-effects, e.g. like largely augmented delay. Therefore, we also study the impact of shorter buffers than necessary on the performance criteria: link utilization, fairness between flows. We also derive a formula for calculating the buffer fluctuation period, in particular Drop-tail buffer synchronization conditions. We show that Drop-tail buffers very often fail to meet the performance criteria. Further on, we present a methodology for optimally dimensioning RED-managed buffers. This is a compilation of other acknowledged works in the area, passed through the prism of our own research and we demonstrate that such buffers, if properly configured and working in a stable regime, can outperform Drop-tail buffers.