Core Concepts#

This section covers fundamental architectural principles and operational behaviors of the NvGrid system for grid integration.

Architectural Principles#

Standards-Based APIs#

NvGrid integrations use standardized gRPC APIs with protobuf message definitions. This approach provides language-agnostic integration, strong typing, and efficient serialization for communication between grid solutions and DPS.

Event-Driven Architecture#

Grid solutions receive real-time power event notifications through webhook systems. This enables immediate awareness of power state changes without polling, supporting responsive integration with utility grid operations.

Best Effort Power Management#

DPS manages power constraints without terminating running workloads. The system reduces power from Dynamic Power Management (DPM) enabled jobs while maintaining operational continuity, pursuing power targets without disrupting datacenter operations.

Multi-Feed Architecture#

NvGrid supports granular control over individual power feeds or entire datacenters. Load targets can be applied to specific electrical circuits or across all feeds, providing flexible power management aligned with utility infrastructure.

Simulator-First Development#

Grid integrations are developed and tested using DPS simulators. This provides consistent testing environments independent of actual utility connections, enabling development and validation without requiring live grid access.

Power Management Strategy#

The NvGrid system implements a best effort strategy for power constraint enforcement.

Constraint Application#

DPM-Enabled Jobs

Power reduction targets Dynamic Power Management (DPM) enabled workloads first. These workloads can have their power limits adjusted without termination.

Non-Terminating Approach

Running jobs are never terminated to meet power constraints. The system prioritizes operational continuity over strict constraint adherence.

Gradual Reduction

Power is reduced incrementally rather than abruptly:

  1. System identifies DPM-enabled workloads with available power headroom

  2. Power limits are reduced in stages across identified workloads

  3. Reduction continues until constraint is satisfied or workload minimums are reached

  4. System monitors actual power consumption throughout the process

Target Achievement Tracking

System continuously tracks whether actual power consumption meets target constraints. Non-achievement is acceptable under the best effort model when constraints require job termination.

Power Restoration#

Automatic Recovery

When power constraints are relaxed or expire, DPS automatically attempts to restore power to DPM-enabled jobs. Power may be distributed to different workloads than those originally reduced.

Power Distribution

Power restoration follows DPS power management policies:

  • DPM-enabled jobs that had power reduced receive increased allocations

  • Power distribution respects workload power policies and limits

  • Restoration occurs incrementally to maintain system stability

  • Topology and resource group constraints govern power allocation

Restoring Default Schedule

To explicitly return to default constraints before a scheduled target expires, schedule a zero watt target for the time period. This cancels any active targets in that time range, allowing the system to revert to default constraints.

Feed Management#

Datacenters are powered by one or more electrical feeds, each representing a distinct circuit. NvGrid provides independent control over each feed.

Power Feeds#

Each feed has operational boundaries defined by metadata:

  • Feed Tag: Unique identifier (e.g., “feed-a”)

  • Power Limits: Minimum and maximum power caps

  • Default Constraint: Power limit when no targets are active

Targeting and Control#

Load targets can be applied to specific feeds or all feeds:

  • Specific targeting: Apply constraints to selected feeds only

  • Global targeting: Apply constraints across all feeds

  • Independent operation: Each feed operates according to its active target

For overlapping time periods, the most recently scheduled load target takes precedence.

Next Steps#

With understanding of these core concepts, proceed to: