Energy-Aware Job Profiling and Mediation for Cloud and 6G Orchestration

Large-scale AI and cloud workloads are driving up data centre energy demand, motivating orchestration that is explicitly energy-aware. This paper presents a practical methodology to characterize and predict job-level energy usage by linking publicly released Google traces: the 2019 Borg workload traces (per-instance CPU/memory usage and job metadata) and co-timed PowerData traces (five-minute power utilization at the power-distribution-unit level). We filter machines by power domain, aggregate per-instance usage into aligned five-minute windows, and construct job-level energy proxies that combine average CPU intensity, runtime, and coincident PDU utilization to derive low/medium/high energy classes. We then train lightweight classifiers (Random Forest, XGBoost) to predict a job's energy class from pre-scheduling usage statistics and runtime, providing a mediation signal that an orchestrator can use to stagger high-energy jobs and backfill with low-energy ones to reduce peaks. The approach delivers two building blocks for sustainable orchestration aligned with 6G systems: empirical energy-job profiling and predictive mediation, grounded in real traces and implementable with modest complexity.