Documentation

Joulie is a Kubernetes-native energy management system that uses per-node digital twins to optimize data center power consumption. It ingests real-time telemetry from every node (CPU/GPU power draw, thermal state, per-pod utilization) to maintain a continuously updated model of the cluster’s energy state. That model drives two things: power cap enforcement (via RAPL and NVML) and scheduling decisions that steer workloads toward the most energy-efficient nodes.

If you are completely new, the smoothest path is:

1. Getting Started
2. Architecture
3. Hardware
4. Simulator
5. Experiments

Core mental model:

• telemetry feeds the digital twin,
• the twin drives operator decisions (power caps, migration triggers),
• the scheduler extender reads twin state to steer new pod placement,
• feedback from new placements updates telemetry, closing the loop.

Section guide

• Getting Started
  • concepts, install, workload compatibility, WorkloadProfile classification, configuration reference
• Architecture
  • operator/agent/twin/scheduler roles, CRDs, policy algorithms, telemetry/control interfaces, kubectl plugin
• Hardware
  • CPU and GPU support model, heterogeneity strategy, runtime caveats
• Simulator
  • trace-driven workload simulation, power modeling, facility stress
• Experiments
  • benchmark design and measured outcomes

What to expect

• Per-node digital twins: telemetry → twin state → cap decisions and scheduling.
• Kubernetes-native contracts: 2 user-facing CRDs (NodeHardware, NodeTwin) + scheduling constraints as intent/supply language.
• Workload classification: automatic profiling via Kepler/cAdvisor metrics with transparent classification reasons.
• Observability tooling: kubectl joulie plugin, Grafana dashboard, Prometheus metrics.
• Practical path to adoption: quickstart first, then progressive deep dives.