Can we trust profiling results?: understanding and fixing the inaccuracy in modern profilers

Profilers are an indispensable component in modern software stack of data centers and supercomputers. Profilers collect detailed performance data during program execution and guide code optimization across the entire software stack. The accuracy of the profiling result proves to be vital for one to effectively gain performance insights. Unfortunately, inaccuracy may arise due to measurement techniques or hardware limits, which can waste optimization efforts.

However, there are few studies in evaluating the accuracy of modern profiling techniques. In this paper, we study performance monitoring units (PMU) based statistical sampling, one of the profiling techniques widely adopted by many state-of-the-art profilers. While PMU sampling based profilers are efficient in collecting performance data, they suffer from inaccurate instruction measurement due to the intrinsic limit in the PMU design. To understand and fix the instruction profiling inaccuracy, we propose a novel 3-step approach. First, we investigate multiple modern architectures and quantify the PMU instruction profiling inaccuracy in these architectures with mathematical modeling. Second, we design a systematic framework to evaluate the impact of PMU inaccuracy to the profiling results. Finally, we propose a software-based technique to rectify the measurement inaccuracy raised by PMU and demonstrate its effectiveness.