Answers are grounded in this paper and cite their sources.
Slides of the talk presenting the paper "Saving Energy Costs and Time on Supercomputers: Scheduler-Integrated Recurring Node Shutdown Exploiting Malleable Jobs" at the 3rd International Workshop on Energy Efficiency with Sustainable Performance (EESP), held in conjunction with SC26 in Chicago, IL, USA, on November 15, 2026. The talk argues that supercomputers can shut down part of their nodes at night, when electricity is more expensive and dirtier than during the day, without making users wait longer, if the job scheduler coordinates the shutdown and malleable jobs release the nodes in time. It covers: - Why nodes sit idle in production supercomputers and why idle nodes still draw about half of their full-load power- Rigid versus malleable jobs, and how malleable jobs fill scheduling gaps- Flex-pool shutdown: a share of the nodes (0, 10, or 30%) is switched off in a recurring off-window from 10 p.m. to 6 a.m., with admission control before and shrink-and-expand of malleable jobs around the window- EASY Backfilling (EBF) versus the malleable Min Steal Agreement (MSA) scheduler under the same shutdown policy- 270 trace-driven simulations on job sets from Cori Haswell, Cori KNL, and Theta with 0 to 100% malleable jobs- MSA reduces flex-pool node overrun by up to 65.2% and keeps on-window utilization above 95%- With 30% of the nodes off at night and only 25% of the jobs malleable, electricity cost and turnaround time both improve on all job sets- Why more malleable jobs are not always better- Next steps: realistic malleable applications, dynamic prices and off-windows, I/O and accelerators Paper:- P. Schäfer, A.-L. Roth, P. Arzt, F. Krull, S. Krenz, F. Wolf, J. Posner: Saving Energy Costs and Time on Supercomputers: Scheduler-Integrated Recurring Node Shutdown Exploiting Malleable Jobs. In: Proceedings of the SC '26 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis (EESP), IEEE, 2026, pp. 1530–1544. doi 10.1109/SCW72768.2026.00195 Software:- ElastisimFulda: Scheduler-Integrated Recurring Node Shutdown in ElastiSim (SC26 EESP artifact), v1.0.0. Extends the ElastiSim simulator with node power states, recurring shutdown policies, dynamic runtime prediction, and a native C++ scheduler interface, and includes the scripts and per-run metrics of all 270 simulations. doi 10.5281/zenodo.22846509- https://github.com/FuldaHPC/artifact-sc26-node-shutdown
No takes yet. Share an insight, caveat, or question.
Schäfer et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: