Towards a Predictive Framework for Power Consumption of Jobs in HPC Facilities

As the mainstream computing technology is entering into a post petascale era, the number and complexity of their computational components is on a sharp increase. With the increased pressure to pack more components per rack, the power and system densities are growing. Recently many researchers are focusing on Power Capping to address the power challenges in current and future computing systems. The power capping can be achieved by proactively estimating the power consumption of High Performance Computing (HPC) Jobs. In this study, we present our proposed machine learning framework to predict the power consumption of Lawrence Berkeley National Laboratory (LBNL) National Energy Scientific Computing Center (NERSC) Cori supercomputer workloads. We evaluate our framework using historical data of real production jobs executed on Cori to predict the amount of power required by a given job and to apply the predictions for enabling power capping in power-limited future systems to be commissioned at LBNL or other installation sites.