Quantifying the Economic Potential of Energy-Aware Scheduling in HPC

High-performance computing (HPC) centers face rising energy costs driven by multi-megawatt demand and utility tariffs that combine time-of-use (TOU) energy prices with monthly demand charges. This work quantifies the economic value of tariff-aware job scheduling in a production setting using five months of data (Jan-May 2025) from a U.S. national laboratory. We evaluate two scenarios against the actual metered baseline: (i) an optimal envelope (Stage A) that reshapes aggregate controllable CPU power under an energy-balance constraint to minimize TOU energy and demand charges, and (ii) an offline, constraint-respecting job schedule (Stage B) that reconstructs a feasible schedule for the historical set of CPU jobs under node-capacity limits and billed-peak guards informed by Stage A. Stage A reduces the total electricity bill by 7.97% relative to actual operations, primarily by lowering monthly demand charges by 17.56%. Stage B, which respects indivisible jobs and capacity constraints, still reduces the total bill by 5.73%, driven by a 12.62% reduction in demand charges. These results show that multi-percent reductions in a full utility bill are available from software-only, tariff-aware scheduling, and we release a reproducible methodology and code implementing this evaluation.