Model Bootstrapping for Auto-Diagnosis of Enterprise Systems

Models for fault diagnosis can help reduce the time taken to accurately identify faults, but the complexity of modern enterprise systems means that the process of manually model-building is itself very time-consuming. We study here the relevance of bootstrapping a diagnostic model that can then be manually refined and augmented by domain experts. We present an approach to model construction, developed by analyzing log traces from a real data center. We compare the automatically-bootstrapped model against a manually-constructed reference model for the same problem set in order to measure what amount of the model can be automatically built. An experiment with an Oracle enterprise system shows that approximately 15% of the model, diagnosing 30% of the related issues, can be automatically built.