Combining hardware and software instrumentation to classify program executions

Several research efforts have studied ways to infer properties of software systems from program spectra gathered from the running systems, usually with software-level instrumentation. While these efforts appear to produce accurate classifications, detailed understanding of their costs and potential cost-benefit tradeoffs is lacking. In this work we present a hybrid instrumentation approach which uses hardware performance counters to gather program spectra at very low cost. This underlying data is further augmented with data captured by minimal amounts of software-level instrumentation. We also evaluate this hybrid approach by comparing it to other existing approaches. We conclude that these hybrid spectra can reliably distinguish failed executions from successful executions at a fraction of the runtime overhead cost of using software-based execution data.