Modeling the GPRS network latency with a double Pareto-lognormal or a generalized Beta distribution

Abstract

Taking a newly collected large data set on the TCP connection termination latency in GPRS networks we try to identify the underlying statistical distribution. The data extends the observed latencies to large time scales necessitating a heavy-tail distribution. Many distributions work well for the main body of the data. However, the heavy tail of the distribution benefits from mixing different statistical distributions. We compare several distributions and find that the double Pareto-lognormal distribution and the generalized Beta distribution of the second kind fit the data equally well.