Detecting repetitive program behaviour from hardware data

Detecting repetitive “phases” in program execution is helpful for program understanding, runtime optimization, and for reducing simulation/profiling workload. The nature of the phases that may be found, however, depend on the kinds of programs, as well how programs interact with the underlying hardware. We present a technique to detect long term and variable length repetitive program behaviour by monitoring microarchitecture-level hardware events. Our approach results in high quality repetitive phase detection; we propose quantitative methods of evaluation, and show that our design accurately calculates phases with a 92% “confidence” level. We further validate our design through an implementation and analysis of phase-driven, runtime profiling, showing a reduction of about 50% of the profiling workload while still preserving 94% accuracy in the profiling results. Our work confirms that it is practical to detect high-level phases from lightweight hardware monitoring, and to use such information to improve runtime performance.