Metrics for Representing Performance as Rewards
In Performability Models, Richard J. Boucherie Z
Behavioral decomposition is a classical modeling approach for performability analysis, typically resulting in Markov reward models, in which the states represent system configurations, and the rewards represent performance given the system configuration. In this report we experimentally investigate ways of obtaining rewards from a performance model. We discuss four candidate metrics to determine rewards, differing by the degreein which transient effects are taken into account: steadystate distribution, asymptotic bias, quasi-stationary distribution and sojourn-time distribution. We compare the metrics with respect to obtained accuracy and practical applicability. The primary conclusion drawn from our experiments is that by basing reward computation on the quasi-stationary distribution instead of the traditionally used steady-state distribution, behavioral decomposition can be…