Raleigh, North Carolina, USA (Red Hat, originating team)
#00158
Case study of
#00165 Measure the emissions of compute you run yourself by sampling hardware power and applying local grid carbon intensity
Implementer
Red Hat Emerging Technologies and IBM Research, with contributors including Intel and Weaveworks; a CNCF project
Timeline
Since Jan 1, 2023
Location
Description
Kepler (Kubernetes-based Efficient Power Level Exporter) attributes energy to individual processes, containers, pods, and nodes in orchestrated environments, exporting results as Prometheus metrics. It draws power readings from RAPL (CPU and DRAM), NVML (NVIDIA GPU), ACPI, and Redfish/IPMI, falling back to a ratio-based model when direct readings are unavailable. Accepted into the CNCF Sandbox in 2023. In June 2026, the project re-architected away from eBPF toward read-only /proc and /sys sampling to improve accuracy and reduce adoption friction. Numbers are best used as consistent relative signals rather than exact ground truth, as CPU-only readings can diverge substantially from whole-node meter readings.
Funding
Lessons learned
- Fine-grained kernel telemetry via eBPF traded accuracy and reliability for granularity — the project reversed course toward simpler /proc and /sys sampling after eBPF became the source of most user-reported issues and missed short-lived processes.
- Validate attribution tool output against whole-node hardware meters before trusting absolute figures; CPU-only estimates can diverge substantially from total node power.
- Exporting to Prometheus makes energy a first-class, trackable metric for platform teams and is a concrete prerequisite for actionability.
Sources
3Documented Jul 13, 2026