Energy Efficient Task Pool Scheduler in OmpSs

Abstract

The European Mont-Blanc project aims to build future exascale systems using energy efficient low-power devices. Exascale systems built using low-power devices will require a large number of processors to achieve competitive performance against state-of-the-art supercomputers. The project relies on the OmpSs programming model and its runtime system, in order to handle the complexity of such a massively parallel system.In this study, an alternative scheduling-plugin has been developed to improve the energy efficiency of the OmpSs runtime system. The proposed scheduling policy from the paper 'Process Cruise Control' has been extended for multi-core systems and integrated into the developed scheduling-plugin. The scheduling-plugin improves the energy efficiency by continuously monitoring the workload, in order to identify situations where it would be beneficial to adjust the frequency through dynamic voltage and frequency scaling. The solution has been evaluated on Sandy Bridge-EP with 17 OmpSs application kernels. Energy consumption is measured for the processor package through the Running Average Power Limit interface on Sandy Bridge. The results shows that energy savings can reach up to 30% in memory intensive applications with limited impact on performance.