The health records system is inflexible; the MRI analysis has limited shifting potential (3 hours 15 minutes of slack); the research pipeline has high shifting potential (days of slack)

The health records and MRI analysis are both inflexible; only the research pipeline can be shifted

All three workloads are inflexible because they operate in a healthcare context where delays are unacceptable

All three workloads are flexible because they are batch processes running on cloud infrastructure

The overnight window is always preferable because cost savings can be used to purchase carbon offsets

Carbon and cost optimisation point in opposite directions in this grid; the scheduling system must accept higher cost for lower carbon, higher carbon for lower cost, or find a compromise window

This conflict does not arise in practice because low-carbon periods always coincide with low-cost periods

The organisation should always choose the midday window because carbon savings should override cost considerations

The proactive scheduler runs jobs faster because it reserves more compute resources in advance

The proactive scheduler can see that a clean window is approaching and hold jobs for it, whereas the reactive scheduler can only see the current intensity and may start jobs during a dirty period or defer them without a plan

The proactive scheduler uses more accurate carbon intensity data

Cloud providers do not allow workloads to be moved between regions

Oregon does not have sufficient datacenter capacity to process the job

Hydroelectric power is unreliable and may not be available when the job runs

The energy and time cost of transferring the large dataset across regions may offset the carbon benefit of running in a cleaner region

Coal plants cannot adjust their output in response to changes in computing demand

The carbon intensity variation across the day is too small to create a meaningful difference between the best and worst execution windows

Coal-dominated grids do not support batch processing workloads

Temporal shifting requires wind or solar generation to function

It forecasts future carbon intensity values for the grid

It classifies incoming workloads as flexible or inflexible based on their resource requirements

It measures the actual energy consumed by each job after execution

It holds deferred flexible jobs until a lower-intensity execution window becomes available, then releases them

Industrial electricity demand drops at midday as factories shut down for lunch breaks

Grid operators deliberately reduce carbon intensity at midday to meet regulatory targets

Solar generation floods the grid during peak sunlight hours, displacing fossil fuel generators and lowering the overall carbon intensity

Wind generation peaks during midday as thermal air currents are strongest

Web search, because it consumes more total energy over the day

Both workloads equally, because they consume electricity from the same grid

Machine learning training, because it tolerates delay and has no live user waiting for an immediate response

Neither workload, because temporal shifting only applies to workloads that can be moved to a different geographic region

Lifecycle carbon intensity

Marginal carbon intensity

Location-based carbon intensity

Average carbon intensity

The mass of greenhouse gas emissions produced per unit of electrical energy generated on the grid

The cost of electricity production adjusted for the type of fuel used

The total amount of electricity consumed by all devices connected to the grid

The efficiency of power distribution infrastructure between the generator and the consumer