Heuristic Formulas to Dimension Buffers in a DSL Access Network Carrying Persistent TCP-controlled Traffic

The capacity of the (DSL) access network keeps increasing, while at the same time there is a trend to support, on top of web browsing applications that spurred the initial popularity of the Internet, more and more applications (e.g. peer-to-peer, time-shifted TV) that require the transport of long-lived flows over such an access network. During the transport of these long-lived flows associated with these (non-timecritical) applications it is important that the last mile link is efficiently used and that each flow gets its fair share of the available capacity and this is what TCP, ideally suited for these non-time-critical applications, aims to do. We work under the assumption that the rest of the network and the upstream, which needs to transport the ACKs, are appropriately dimensioned and thus we concentrate on the downlink and its associated link buffer (in a DSL access network it is in the DSLAM) which has a crucial impact on the efficiency and the fairness of the transport in “last mile”. Studies show that persistent TCP flows have the greatest impact on buffer requirement (and not short-lived (web) flows). We show that both requirements can be fulfilled by choosing the buffer large enough (so that there is no packet loss). We develop a heuristic formula to calculate the required buffer size. Unfortunately, this buffer size may turn out to be too large to be implemented in state-of-the-art access equipment. So, we also show how efficiency and fairness react to a smaller buffer in case the buffer acceptance discipline is Drop-tail or RED. We show that Drop-tail cannot achieve fairness and that RED, if appropriately dimensioned, can (and we also present a methodology for its configuration).