Hardware Counters' Space Reduction for Code Region Characterization.

This work proposes that parallel code regions in an OpenMP application can be characterized using a signature composed by the values of a set of hardware performance counters. Our proposal is aimed towards dynamic tuning and, consequently, the metrics must be collected at execution time, which limits the number of metrics that can be measured. Therefore, our main contribution is the definition of a methodology to determine a reduced set of hardware performance counters that can be measured at application’s execution time and that still contains enough information to characterize a parallel region. The proposed methodology is based on principal component analysis and linear correlation analysis. Preliminary results show that it can be used to successfully reduce the number of hardware counters needed to characterize a parallel region, and that this set of counters can be measured at run time with high accuracy and low overhead using counter multiplexing.