Metrics for Representing Performance as Rewards in Performability Models

Behavioral decomposition is a classical modeling ap proach for performability analysis typically resulting in Markov reward models in which the states represent system con gurations and the rewards represent per formance given the system con guration In this report we experimentally investigate ways of obtaining rewards from a performance model We discuss four candidate metrics to determine rewards di ering by the degree in which transient e ects are taken into account steady state distribution asymptotic bias quasi stationary dis tribution and sojourn time distribution We compare the metrics with respect to obtained accuracy and prac tical applicability The primary conclusion drawn from our experiments is that by basing reward computation on the quasi stationary distribution instead of the tra ditionally used steady state distribution behavioral de composition can be applied to a considerably larger class of models without loss of accuracy