A new approach to detecting execution phases using performance monitoring counters

Abstract

In this paper, a new hierarchical view of the workload phase classification problem is introduced. Execution phases are the continuous pieces of execution that show consistent behaviour in terms of performance and power. To the best of our knowledge, this is the first work which uses a hierarchical approach to collect and cluster the performance monitoring counters in order to detect macroscopic phases in an application. Our results show the ability of our model to differentiate between execution phases according to the processor power behaviour. Furthermore, we investigate the power consistency inside each phase. The results show the effectiveness of our proposed methodology in classifying phases with similar power behaviour. This information can be used by the system to control and maintain power bursts, increasing the data centre’s power efficiency by reducing the maximum-to-average power ratio.