Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices
Authors: Paul Schäfer and Anna-Lena Roth (Fulda University of Applied Sciences); Peter Arzt (Technical University of Darmstadt); Felix Krull and Sebastian Krenz (Fulda University of Applied Sciences); Felix Wolf (Technical University of Darmstadt); and Jonas Posner (Fulda University of Applied Sciences, Germany)
Abstract: Supercomputers consume electricity around the clock, yet photovoltaic generation can make power cheaper and cleaner by day than at night. Thus, shutting down idle nodes overnight could reduce energy cost but risks longer job waiting times. This paper studies scheduler-integrated shutdown policies that proactively power down a share of the nodes during recurring off periods, such as nights. To compensate, we exploit malleable jobs, which the scheduler shrinks before shutdowns to release nodes and expands afterward. In 270 simulations on job sets from three production supercomputers, we evaluate shutdown shares up to 30% under rigid and malleable scheduling with varying fractions of malleable jobs. Shutting down 30% of the nodes benefits both operators and users once 25% of the jobs are malleable. Compared with rigid scheduling without shutdown, operators save up to 12.5% of the daily electricity cost and users up to 56.2% of their job turnaround time.
Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing
Back to Full Workshop Archive Listing