Bottleneck Footprints in TCP over Mobile Internet Accesses

In this paper we propose a novel method to detect a bottleneck in the core section of a mobile network using TCP related counters. The dataset used in this work consists of the full TCP statistics for two different UMTS SGSNs over four different one-day periods: one day in March 2006 and three consecutive days in September 2006. We derived the number of packets and the number of retransmissions per individual user during the peak hours. These two quantities were put into scatterplots to derive "footprints" of global TCP behavior in the SGSN areas. One of the traces contained a capacity bottleneck on an SGSN link. This trace was taken as the reference footprint for bottleneck presence. Based on such reference we developed a method to infer the presence of future bottlenecks based on footprint similarity. We tested two different distance metrics. It turns out that the simple correlation performs similar to a much more sophisticated function based on a ratio of symmetriezed Kullback-Leibler distances.