Power Compliance#
DPS delivers modeled power compliance by construction. Before it changes managed power, it evaluates the topology, effective power policies, configured limits, and known load. When a change transfers capacity, DPS reduces power first and permits an increase only after it confirms the reduction. If it cannot confirm the reduction, it keeps the existing or conservative reservation.
This reduce-then-increase behavior keeps an in-progress change within the modeled boundary, even when a device response is delayed or incomplete. When you give a workload more power, you also change demand on a shared PDU, busbar, or facility feeder. A qualified meter at that same boundary establishes the physical outcome.
This guide applies the distinction in Provisioned and Observed Power to a declared electrical boundary. It explains how DPS keeps the modeled constraint intact while a power change is in flight, and how boundary metering confirms the result.
Important
DPS can establish what it planned and what controllable devices acknowledged. A fresh, qualified meter reading at the declared boundary establishes what the electrical system actually did.
Begin With the Boundary#
Start with the facility constraint that you need to protect. The boundary can be a facility feeder, rack PDU input, DC busbar, compute-tray input, or a controllable device. It defines which loads, limits, and measurements belong in the same decision.
Use that boundary consistently. The operating limit, provisioning margin, contributing loads, and meter must describe the same electrical path. A device-level cap cannot, on its own, establish the state of a facility-level boundary. Equipment in one rack can also be supplied by different feeds or busbars, so physical location is not enough.
Build an Honest Capacity Model#
With the boundary defined, DPS combines topology, effective policies, configured limits, and static loads to determine modeled capacity. This model turns a facility constraint into a workload decision.
Power policies constrain what controllable devices can request. They do not represent every load behind a tray, rack, or facility boundary. You must model, measure, or reserve every material load at the selected boundary.
Use the StaticLoad field on an entity for modeled fixed
consumption. Treat it as a conservative engineering allowance, not as an
observed meter reading. Keep unknown or uncontrolled load reserved until you
can measure it at the boundary.
Change Power Without Reusing Capacity#
Consider a resource group that needs more GPU power while another resource group can give some back. DPS reduces the donating workload before it makes that capacity available to the receiving workload. This order prevents both workloads from using the same modeled capacity during a transition.
If DPS cannot confirm the reduction, it keeps the prior or conservative reservation. The model remains aligned with the state that the control plane can substantiate, even when a device response is delayed or incomplete. DPS Power Steering Controller follows the same model when it proposes per-GPU updates, and DPS validates those updates before it applies them to hardware.
Use Metering to See the Physical Outcome#
The modeled result is an expectation. Metering tells you what happened at the electrical boundary. For a measurement-compliance claim, your meter must observe the same physical boundary as the configured limit.
Telemetry and observability data help you monitor and optimize devices. They do not automatically establish power at a rack, PDU, or facility boundary. To make a meaningful claim, the meter must identify the boundary, units, time basis, quality, and aggregation. Treat missing, stale, implausible, or unscoped readings as unknown — not as zero power.
In deployments that use Zapp, the Telemetry Provider streams sensor data from datacenter entities into DPS, which uses that real-time power data for excursion detection and mitigation. It complements, rather than replaces, the qualified boundary meter required for a measurement-compliance claim. Refer to Configure Zapp Telemetry to configure the integration.
Compare the model and the boundary meter after commissioning and material changes. Differences can reveal unaccounted load, a measurement mismatch, conversion loss, or a changed operating state.
Apply the Right Time Horizon#
Your facility can protect the same boundary at more than one timescale. Keep those limits distinct.
50 ms defines a short-duration or peak-power envelope. DPS does not establish a 50 ms measurement-compliance result. If your site needs this claim, use a qualified facility measurement and analysis path at the declared boundary.
1 second is the primary continuous power window. Boundary metering at this timescale establishes whether the continuous operating limit is met.
When you configure the
simpleexcursion mitigation strategy, its default telemetry wait after a reduction is 10 seconds. It gives DPS time to obtain fresh telemetry; it is not an allowable duration above the 50 ms or 1-second limit.
Define Excursions With the Facility#
Do not infer an allowable excursion from a DPS policy, allocation result, or mitigation timer. An allowable excursion is a facility-approved short-duration envelope. The facility design defines its boundary, limit, time window, measurement method, and required response.
DPS does not create a permitted excursion duration or count. Without an explicitly approved envelope, an over-limit result is non-compliant at the applicable 50 ms or 1-second timescale. The 10-second mitigation interval does not change that result.