Model-based Fault Localization: Finding Behavioral Outliers in Large-scale Computing Systems

We present a model-based approach to fault localization that aims to help the human analyst narrow down the manual localization into a small fraction of the overall system. Our method consists of two parts: pre-failure model derivation and post-failure model-based anomaly detection. The first part collects function-call traces from all processes and derives an execution model that reflects the function-calling behaviors of the target system. When a failure occurs, we identify the most deviant behaviors in the failed run by comparing the failure traces with the derived model. We claim that the analyst can substantially reduce the burden of fault localization by prioritizing such behaviors. Our preliminary experiment with a distributed job manager supports this claim: Our method narrows down localization of a 70-second faulty run on a 78-node distributed platform into just sub-second behaviors involving only two nodes.