Granular Kernel-Level Telemetry Drives Carbon-Aware Scheduling Across Kubernetes Clusters

The Kepler (Kubernetes-based Efficient Power Level Exporter) project provides kernel-level energy telemetry for cloud-native infrastructure, measuring dynamic wattage across containers, pods, and nodes by combining extended Berkeley Packet Filters (eBPF) with hardware interfaces such as Intel Running Average Power Limit (RAPL) and ACPI. By continuously attributing power consumption to specific Linux control groups (cgroups) and exporting standard Prometheus metrics, the platform bridges the gap between low-level hardware counters and application-level orchestrators. For platform and DevOps engineers, this shifts sustainability from an abstract corporate compliance exercise into an engineering discipline. Historically, cloud carbon tracking relied on post-facto estimation models provided in hyperscaler billing reports. These macro-level figures lack the granularity required to answer operational questions, such as whether a specific microservice revision increased CPU package wattage or which background processing queue consumes the most power per transaction. By surfacing container-level energy consumption alongside standard CPU and memory utilization, teams can pinpoint power anomalies, detect inefficient software loops, and accurately assess the environmental impact of distributed workloads. This tooling arrives at a pivotal moment for cloud operations. Escalating power requirements from generative AI pipelines and data-intensive processing have forced data center operators and enterprises to reevaluate power efficiency thresholds. While cloud providers continue investing in renewable energy procurement, software architectures must evolve toward carbon-aware computing—modulating compute demand based on grid intensity and energy efficiency curves. Kepler acts as the telemetry foundation required by higher-level tools, such as custom Kubernetes schedulers and KEDA scalers, to execute carbon-aware time-shifting and location-shifting policies. In practice, infrastructure teams should evaluate Kepler by deploying the operator into non-production Kubernetes environments and integrating its metrics endpoints with existing Prometheus and Grafana dashboards. On bare-metal or on-premises nodes, Kepler extracts direct RAPL energy readings; in multi-tenant public cloud virtual machines where direct access to hardware counters is restricted, it uses machine-learning estimation models calibrated against baseline hardware configurations. SREs can initially use this data to identify idle workloads and optimize pod resource limits. Over time, teams can feed energy metrics into CI/CD performance testing and custom scheduling algorithms to ensure that compute-heavy batch tasks run only when power efficiency is maximized.