Speaker
Description
Energy efficiency is a critical concern for WLCG operations. We present a proof-of-concept for dynamic power accounting in our heterogeneous compute clusters at ScotGrid Glasgow. Our approach leverages real-time metrics from Prometheus to attribute energy consumption to individual Virtual Organizations (VOs) based on actual core usage. By integrating hardware-specific power efficiency data, derived from static measurements across different node generations, we compute per-core power usage while accounting for architectural differences.
Our methodology distinguishes between active power (consumed by running jobs) and infrastructure overhead (idle power and other services), the latter is allocated to the hosting institute. This granular, data-driven model not only provides transparent energy allocation but also encourages system administrators to optimize resource utilization and improve overall Power Usage Effectiveness (PUE).
Our work lays the foundation for integrating energy accounting into existing monitoring infrastructures and provides insights into sustainable cluster operations.
| Requested talk length | 20 |
|---|
