Speaker
Description
The increasing power densities of exascale supercomputers require complex cooling infrastructures. ExaDigiT [1] addresses this need by providing a digital twin for liquid-cooled supercomputers through Functional Mock-up Unit (FMU)-based thermo-fluid models. These FMU simulators are accurate but computationally expensive for rapid what-if scenario exploration and Monte Carlo uncertainty quantification studies. Hence, replacing these simulators with efficient deep learning surrogates is a key step toward energy-aware digital twins. But effective surrogate design requires a prior understanding of the input-output relationships, temporal dynamics, non-linearities, and spatial coupling. This work addresses that gap through a systematic statistical analysis of FMU-generated data for three IBM POWER9 systems: Marconi100, Summit, and Lassen, all within the ExaDigiT framework. We performed six progressive analyses: (i) direct effect, (ii) multivariate effect, (iii) rate of change, (iv) non-linearity, (v) spatial interaction, and (vi) physics consistency.
Our key findings are (i) a structural decoupling between the primary and secondary loops, as the primary loop is actively managed and substantially nonlinear and load-following; (ii) level-driven dynamics (median rate/level ratio = 0.33) with 54% of pathways exhibiting higher-order transient responses and Autocorrelation Function (ACF) persistence beyond 1,000 lags; and (iii) negligible inter-CDU thermal propagation that was confirmed by distance-independent correlations and asymmetry ratio Rasym = 0.95. and (iv) near-zero thermodynamic violations with tolerance-calibrated energy conservation constraints.
REFERENCES
[1] Wesley Brewer et al. “A Digital Twin Framework for Liquid-cooled Supercomputers as Demonstrated at Exascale”. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis. SC ’24. Atlanta, GA, USA: IEEE Press, 2024. DOI: 10.1109/SC41406.2024.00029. URL: https: //doi.org/10.1109/SC41406.2024.00029.