Adaptive Resource and Job Management for Limited Power Consumption

International audienceThe last decades have been characterized by anever growing requirement in terms of computing and storage resources.This tendency has recently put the pressure on the abilityto efficiently manage the power required to operate the hugeamount of electrical components associated with state-of-the-arthigh performance computing systems. The power consumption ofa supercomputer needs to be adjusted based on varying powerbudget or electricity availabilities. As a consequence, Resourceand Job Management Systems have to be adequately adaptedin order to efficiently schedule jobs with optimized performancewhile limiting power usage whenever needed.We introduce in this paper a new scheduling strategy thatcan adapt the executed workload to a limited power budget. Theoriginality of this approach relies upon a combination of speedscaling and node shutdown techniques for power reductions. It isimplemented into the widely used resource and job managementsystem SLURM. Finally, it is validated through large scale emulationsusing real production workload traces of the supercomputerCurie