An approach for classifying program failures

In this work, we leverage hardware performance counters-collected data to automatically group program failures that stem from closely related causes into clusters, which can in turn help developers prioritize failures as well as diagnose their causes. Hardware counters have been used for performance analysis of software systems in the past. By contrast, in this paper they are used as abstraction mechanisms for program executions. The results of our feasibility studies conducted on two widely-used applications suggest that hardware counters-collected data can be used to reliably classify failures.