Power-Aware & Grid-Responsive Orchestration

Glossary

Data gravity

The pull that large datasets exert on compute, because moving or replicating big data across regions is costly, slow, and sometimes legally restricted. It sets a job's freedom to be relocated.

Deferrable workload

Compute with slack in its deadline, such as batch inference, evaluation sweeps, or training, that can be delayed or relocated without breaking a service promise. It is the raw material for all workload shifting.

Demand response

Programs through which a grid operator pays consumers to reduce or shift demand when the grid needs it, turning a data center's orchestration flexibility into capacity and energy revenue, priced by response speed and reliability.

Grid-interactive load

An electricity consumer that can modulate its demand in response to grid signals on a timescale the grid operator can use, substituting flexible load for reserve generation during stress.

Marginal carbon intensity

The emissions of the specific generator that responds to one additional unit of demand at a given time and place, the correct signal for a load deciding whether to draw now or later, as opposed to the whole-grid average.

Power-aware orchestration

Scheduling that treats electricity cost, carbon intensity, and proximity to grid limits as first-class inputs to where and when a job runs, on equal footing with hardware availability, bounded by each job's deadline.

Spatial shifting

Routing a deferrable job to a different region whose grid is currently cheaper, cleaner, or less congested, exploiting differences across sites at one instant, subject to data-movement cost and latency limits.

Temporal shifting

Moving a deferrable job to a more favorable hour on the same hardware, exploiting daily swings in power price and carbon intensity while keeping the work in place. The lowest-risk shifting technique.