The Performance-Power Frontier: A Model-Driven Approach to Energy-Aware Application Optimisation

Energy consumption in high-performance computing (HPC) has emerged as a primary design constraint, with Tier-0 systems now demanding tens of megawatts. While users are increasingly being encouraged to minimise the environmental and operational costs of their workloads, they frequently lack the analytical frameworks required to navigate the complex trade-offs between execution time and energy efficiency. In this paper, we extend the Power-Optimised Software Envelope (POSE) model and introduce a visualisation tool designed to help performance engineers explore the energy-time design space. By aligning multi-objective optimisation metrics with institutional operational priorities, our framework enables policy-aware decision-making at the application level. We demonstrate this approach by analysing the NAS Parallel Benchmarks. Our results reveal a significant variance in energy-aware optimisation potential: while the Embarrassingly Parallel (EP) kernel shows negligible opportunity for gains (up to 0.37% improvement in Energy-Delay Summation), the Conjugate Gradient (CG) kernel offers greater scope (up to a 12.96% improvement). Furthermore, we study the benchmarks under frequency and power constraints and identify critical efficiency thresholds; for the Block Tri-Diagonal (BT) solver, configurations operating below 2.00 GHz or under a 140 W powercap fail to outperform the baseline efficiency, highlighting the risks of aggressive over-throttling.