Compare feasible schedules with all declared movement costs

Radovanović and colleagues describe carbon-aware capacity shaping for temporally flexible work using demand and carbon forecasts. The key engineering idea is a constrained schedule: work must still complete within its allowed horizon and capacity. Interactive requests with immediate response requirements have different flexibility from a batch analysis due tomorrow. State release times, dependencies, duration estimates, and deadline slack before deciding that a future interval is actually available.

Spatial movement adds data placement, allowed regions, network capacity, and destination availability. Temporal movement can accumulate a queue that later competes with other delayed work. Forecasts of both electricity factors and execution time can be wrong. The original workload-shifting limitations research examines why flexibility assumptions matter; its results do not establish a universal achievable reduction. Evaluate fallback schedules and deadline risk rather than optimizing a forecasted scalar while ignoring feasibility.

Compute a shifted account from disjoint execution, transfer, and storage energy components, each with the factor matching its location and interval. A slower destination can use more execution energy even if its factor is lower. A preexisting copy may avoid some transfer, whereas repeated checkpoints or retention can add costs. The boundary must say whether each energy term is IT-only or facility-inclusive; do not apply a second facility multiplier to an already inclusive component.

Separate an average-factor inventory scenario from a marginal-impact scenario. Forecast average factors can rank a declared accounting objective, while a consequential claim needs an appropriate response model and horizon. Compare predicted and realized inputs afterward using the same workload and factor conventions. If movement costs or uncertainty reverse the sign, report that possibility. A numerical optimization can be internally correct while its omitted components make the headline claim unsupported.

A scheduling improvement for fixed work does not determine the effect of changing demand. Lower cost or improved capacity can allow more jobs to run. For example, multiplying per-job energy by a larger job count can increase the total even when the per-job figure falls. This is a conditional rebound scenario, not a claim that every efficiency improvement inevitably increases consumption. Measure actual workload volume and report absolute totals alongside normalized ratios.

Keep a reversible scheduling policy and an auditable record of why each shift was chosen: forecast version, candidate region or interval, work estimate, transfer assumptions, and deadline margin. Monitor completed work, failures, and additional storage rather than evaluating only jobs that successfully shifted. A policy that drops difficult work can look efficient by selection. The decision should remain inspectable when an updated factor, duration estimate, or demand forecast changes the preferred schedule.