PH and MAP Fitting with Aggregated Traffic Traces

Abstract

Phase Type Distributions (PHDs) and Markovian Arrival Processes (MAPs) are versatile models for the modeling of timing behavior in stochastic models. The parameterization of the models according to measured traces is often done using Expectation Maximization (EM) algorithms, which have long runtimes when applied to realistic datasets. In this paper, new versions of EM algorithms are presented that use only an aggregated version of the trace. Experiments show that these realizations of EM algorithms are much more efficient than available EM algorithms working on the complete trace and the fitting quality remains more or less the same.