Workshops Archive

Variation-Aware Non-Uniform Node Power Capping for Exascale HPC Systems


Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices

Authors: Solomon Bekele (Argonne National Laboratory); Brad Geltz (Intel Corporation); Swann Perarnau (Argonne National Laboratory); Siddhartha Jana and Christopher M. Cantalupo (Intel Corporation); and Servesh Muralidharan, Kevin Harms, Paul Rich, Brice Videau, and Kalyan Kumaran (Argonne National Laboratory)

Abstract: Exascale systems such as Aurora operate under stringent power and thermal constraints, yet their nominally identical nodes exhibit substantial performance-per-watt variation due to manufacturing, aging, and thermal differences. Under uniform node power caps, this variation slows the critical-path node in tightly coupled applications and reduces the effectiveness of a fixed job-level power budget. We present a variation-aware, non-uniform node power allocation method for Aurora that (i) characterizes the power–performance response of individual nodes using a large-scale RAPL (Running Average Power Limit) power sweep, (ii) represents the characterization using either a quadratic or a piecewise-linear model, and (iii) uses the selected model to compute non-uniform per-node power caps under a fixed job-level power budget. Using measured Nekbone experiments at approximately 512 nodes and 8,210 nodes, non-uniform allocation improves normalized application performance by up to 7.5% at moderate scale and 6.8% at 8,210 nodes compared with matched uniform node power capping.


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