Towards Sustainable HPC Operations: Job-Level Modeling of Power and Carbon Emissions

High-Performance Computing (HPC) systems are critical to scientific discovery and large-scale data processing, yet their growing energy demand raises urgent sustainability challenges. This thesis presents a methodology to model job-level power consumption and compute corresponding operational carbon footprints by combining power models with time- and location-specific carbon intensity (CI) forecasts. Using the CINECA Marconi100 dataset, node- and job-level power models (linear and second-order polynomial regressions) are fit with inputs including GPU, CPU, memory utilization, fan speeds, and ambient temperature; node-level models achieve 6-10% Mean Absolute Percentage Error (MAPE) while job-level models reach 13% MAPE. The analysis of metrics also revealed ambient temperature gradient across chassis correlating with node power consumption. For CI prediction, a Prophet time-series model at 30-second resolution attains 11.3% MAPE. Integrating per-job power estimates with CI forecasts yields per-job CO2 estimates that closely align with measured baselines (based on the real measured power and historical CI) over a ten-day trace (within 1% of the baseline). To illustrate how models could enable carbon-aware decisions, proof-of-concept scheduling experiments were performed on a single day job data, the highest number of jobs in the observation period. Two strategies were evaluated: (i) temporal rescheduling of flexible jobs to predicted low-CI windows (100% and 75% reschedulable scenarios), and (ii) temporal rescheduling combined with placement onto cooler (lower-chassis) nodes (simulated by adjusting ambient temperature for moved jobs). Using both the proposed per-job power and CI models and a measured power with historical CI baseline, modest emission reductions were observed (3.1-3.6%). These experiments are proof-of-concept: they demonstrate the potential for carbon-aware policies but do not quantify production costs, which are left to future work.