This document discusses using ergodic hidden Markov models to characterize workloads. It describes workload characterization as creating a model from measured workload data like log files or traces. The approach treats sequences of virtual pages as time-varying data and analyzes them with statistical techniques and hidden Markov models. The models can determine the type of workload and generate similar log files. The document outlines parameters for the hidden Markov model like using short-time spectral analysis on page references and defining a spectral distance metric between logs. It evaluates using discrete and continuous hidden Markov models to classify single traces and multiple traces of the same workload to model program behavior. The conclusion states this approach achieved a 76% classification accuracy rate.