This paper studies the use of statistical induction techniques as abasis for automated performance diagnosis and performance management.The goal of the work is to develop and evaluate tools for offline andonline analysis of system metrics gathered from instrumentation inInternet server platforms. We use a promising class of probabilisticmodels (Tree-Augmented Bayesian Networks or TANs) to identifycombinations of system-level metrics and threshold values thatcorrelate with high-level performance states--compliance withService Level Objectives (SLOs) for average-case response time--in athree-tier Web service under a variety of conditions.