Mining precise performance-aware behavioral models from existing instrumentation

Software bugs often arise from differences between what developers envision their system does and what that system actually does. When faced with such conceptual inconsistencies, debugging can be very difficult. Inferring and presenting developers with accurate behavioral models of the system implementation can help developers reconcile their view of the system with reality and improve system quality. We present Perfume, a model-inference algorithm that improves on the state of the art by using performance information to differentiate otherwise similar-appearing executions and to remove false positives from the inferred models. Perfume uses a system's runtime execution logs to infer a concise, precise, and predictive finite state machine model that describes both observed executions and executions that have not been observed but that the system can likely generate. Perfume guides the model inference process by mining temporal performance-constrained properties from the logs, ensuring precision of the model's predictions. We describe the model inference process and demonstrate how it improves precision over the state of the art.