Analysis of TCP Bottleneck Footprints in Traces from live 3G Networks

In this paper we evaluate four different metrics for non intrusive bottleneck detection based on TCP counters. Three using a reference and one being reference free. This work is based on the full TCP statistics recorded on five days over the last one and a half year within the live core network of a mobile network operator in Austria. We used two of the datasets, which had a known bottleneck in place as a reference for the metrics. We generated scatterplots, so called footprints, using the number of packets and the number of retransmissions for each user in the peak hours around 7–11 p.m. for all traces. Based on the bottleneck reference we tested the metrics for the detection probability of a bottleneck event. All four metrics detected the bottleneck with no problem, also the throughput figures between 2006 and 2007 had increased. The performance of a simple correlation performance was similar to other metrics based on more sophisticated functions.