Speaker
Description
Monitoring power consumption at the level of grid job slots remains a missing component of current Workload Management Systems for HEP experiments. While individual computing centres can monitor power consumption locally, maintaining a consistent view across heterogeneous clusters and re-benchmarking systems after each configuration change is time-consuming and often impractical for sites.
At the WLCG Workshop last year, we proposed a lightweight approach to address this gap by integrating power measurements into existing WLCG benchmarking and workload infrastructures. Over the past year, this approach has been used in practice through adoption at new sites and systematic data collection, enabling iterative refinement and validation based on real production data. In this contribution, we present the current status of the implementation together with the first results obtained from production environments.
Two power collectors are currently recommended to cover the majority of WLCG sites: a systemd-based collector and an implementation designed for sites already using Prometheus. Both solutions are straightforward to deploy and require minimal effort from site administrators, lowering the barrier for adoption.
Over the past year, the work has evolved from a proposed concept to a validated implementation supported by real production data and cross-site analysis. Initial results allow us to explore performance-per-watt characterisation, anomaly detection, and consistency of measurements across heterogeneous environments. Broader site adoption will enable more robust cross-site comparisons and improved modeling. In turn, this will support carbon footprint estimation per job, representative HS23/Watt values even for sites without direct power measurements, and more informed operational and hardware decisions across the WLCG infrastructure.
| Desired slot length | 20 |
|---|---|
| Speaker release | Yes |