Hardware Lifecycle-Aware Power Planning in Datacenters

This technical report presents a comprehensive study on power planning methodologies for commercial hyperscale datacenters, specifically focusing on the challenges posed by heterogeneous hardware. The authors detail their practical experiences at Meta over the past decade, emphasizing the need for effective power budgeting strategies that account for both legacy and new hardware. The report introduces a hardware lifecycle-aware power budgeting methodology, which has been successfully implemented to achieve an average power oversubscription of approximately 20%. It also discusses the development of PowerSight, a machine learning-based model designed to predict system power consumption during early hardware lifecycle phases when production-level telemetry is unavailable. The paper outlines the importance of understanding power characteristics across different hardware generations and workloads, and it highlights the necessity of adapting power planning approaches as hardware transitions through various lifecycle stages. The findings aim to enhance power management practices in datacenters and contribute to more efficient energy usage.