Introducing Energy based fair-share scheduling

International audienceEnergy consumption has become one of the most important parameters in High Performance computing platforms. Fair-share scheduling is a widely used technique in job schedulers to prioritize jobs, depending to past users allocations. In practice this technique is mainly based on CPU-Time usage. Since power is managed as a new type of resources by Slurm and energy consumption can be charged independently, there is a real need for fairness in terms of energy consumption.This presentation will introduce fair-share scheduling based on past energy usage in Slurm. The new technique will allow users that have optimized their codes to be more energy efficient or make better usage of DVFS techniques to improve the stretch times of their workload