Aiming to Optimize HPC Resource Utilization based on Jobs' Energy Consumption Data

Optimization of energy consumption and HPC resource utilization are high-priority issues for any supercomputing centre . Accelerators will drive the floating-point performances of exascale supercomputers; however preparing the scientific communities for these ever-changing architectures is a daunting task. Thus, the work of HPC support teams is crucial in achieving an overall optimized resource utilization. We highlight a methodology to monitor and optimize resource utilization at a national supercomputing centre, CINES [8] . We used a tool to collect the total energy consumed (kWh) by the simulation jobs on supercomputer ‘Occigen’ for several months. Based on  benchmark runs and our data, we prepared a classification criterion to distinguish jobs based on their normalized energy consumption; and classified them from low to high-energy consumption classes. Our data analysis shows interesting  trends that help us identify suboptimal jobs, along with other insights, to improve resource utilization as a final goal on CINES’s supercomputers