Cloud Telemetry Modeling via Residual Gauss-Markov Random Fields

Can probabilistic graphical models characterize cloud telemetry? This paper promotes the affirmative view. Cloud systems are large, connected, and dynamic. Consequently, databased techniques to model their telemetry are high-dimensional, spatial, and unsupervised. Undirected probabilistic graphical models seem natural, but remain unexplored. We discuss one way around the limitation that cloud measurements violate usual assumptions of normality, and give a tractable estimation procedure for a candidate data model. As a preliminary test, we fit the model and use it to detect and localize anomalies in a synthetic environment and for a small-scale software system.