Estimating performance models parameters of cloudsystems presents several challenges due to the distributed natureof the applications, the chains of interactions of requests witharchitectural nodes, and the parallelism and coordination mech-anisms implemented within these systems.In this work, we present a new inference algorithm for modelparameters, calledstate divergence(SD) algorithm, to accuratelyestimate resource demands in a complex cloud application.Differently from existing approaches, SD attempts to minimizethe divergence between observed and modeled marginal stateprobabilities for individual nodes within an application, thereforerequiring the availability of probabilistic measures from both thesystem and the underpinning model.Validation against a case study using the Apache CassandraNoSQL database and random experiments show that SD can ac-curately predict demands and improve system behavior modelingand prediction.