An experimental study of router buffer sizing for mixed TCP and real-time traffic

Recent research results on Internet router buffer sizing suggest that when TCP traffic is well-paced, Internet routers need as few as 20-30 packet buffers. to realise near-maximum link throughputs, independent of link capacities and number of TCP flows. However, these studies have ignored nonTCP traffic, on the grounds that TCP traffic predominates in the Internet. In this paper we evaluate this assumption via practical experiments on a long-haul Australian network. Specifically, for different volumes of long range dependent real-time streaming video traffic in the network, we quantify end-to-end TCP throughput and real-time traffic loss as router buffer sizes at bottleneck and non-bottleneck links vary. Our results indicate that even in the presence of 5-15% bursty real-time traffic, TCP requires larger router buffers to achieve a given fraction of its saturation throughput. However, larger buffers lead to increased losses for real-time traffic. This suggests that TCP and non-TCP traffic can negatively impact each other, and their performance trade-offs need to be considered when sizing router buffers.