From Modeling to Decision Intelligence: Learning to Navigate Power–Performance Trade-

In this invited talk on navigating power-performance trade-offs, Tania Islam, associate professor at Texas State University, argues that energy efficiency in HPC must be treated as a multi-objective decision-making problem rather than a pure prediction task. She surveys mechanisms such as per-node and facility-level power caps, shifting power between jobs according to their compute- or IO-bound character, and DVFS or power throttling, using HPC scheduling as a running example: launching all ready jobs at once causes power spikes that trigger cooling costs, while schedulers must also balance throughput and fairness. She critiques existing predictive models as black boxes that answer the forward problem (given a configuration, predict time and power) when scheduling actually requires the inverse problem: which actionable configuration achieves a target. Two recent efforts are presented. The first uses job fingerprinting with neurosymbolic AI to produce explainable, auditable scheduling decisions, reducing average wait and turnaround times in a digital-twin scheduler simulation based on an Oak Ridge National Laboratory framework. The second improves actionability of inverse solutions by ranking candidate configurations with uncertainty quantification. Islam closes by asking whether AI-driven system management is itself energy efficient, or whether the community is merely relocating energy costs into the intelligence layer. In the Q&A she discusses the difficulty of deploying such methods on production schedulers and the need to model the power costs of AI components themselves.