CSR---PDOS: Model-Driven Comprehensive Performance Anomaly Characterization for System Software

University Of Rochester, Rochester NY

Investigators

Abstract

Modern operating systems are large, complex software which are developed through collaboration of hundreds of people over years. They may perform anomalously due to various implementation problems, which cause performance degradation, and more importantly compromise the predictability of system performance behaviors. This project investigates a new approach to comprehensively examine the large space of possible runtime settings (including workload properties and system configurations) and to characterize performance anomalies in system software like operating systems. This approach relies on the knowledge of high-level system design algorithms/protocols and employs a system performance prediction model using such knowledge. Aided by the model, the approach acquires a representative set of anomalous workload and system configuration settings by comparing measured system performance with model prediction at some sample settings. Anomalous settings are then statistically clustered into groups likely attributed to individual ``causes''. Finally, each such cause (or performance bug) is characterized with correlated system configurations and workload properties. The research contribution of this project is the proposal of a systematic approach for comprehensively characterizing performance anomalies in complex system software. Such anomaly characterization can aid performance debugging and guide the avoidance of anomaly-inducing runtime settings. Broader impacts of the project include the dissemination of research results and developed software artifacts. In parallel to research, this project also enhances the systems-area curriculum at the University of Rochester, with the emphasis on having students recognize the existence of performance anomalies in complex systems and appreciate the difficulty to understand them.