PowerDis: Fine-Grained Power Monitoring Through Power Disaggregation Model

Xinxin Qi, Juan Chen, Rongyu Deng, Zekai Li, Lin Deng, Yuan Yuan, Yonggang Che

Abstract: In an era where power and energy are the first-class constraints of computing systems, precise power information is crucial for energy efficiency optimization in parallel computing systems. Current power monitoring techniques rely on either software-centric power models that suffer from poor accuracy or hardware measurement schemes that have coarse granularity. This paper introduces PowerDis, a new technique for accurately measuring and forecasting the power consumption of computing components, such as CPUs and memory, within the compute node on parallel computing systems. PowerDis combines coarse-grained node power readings collected through hardware measurement and a software power disaggregation model to improve monitoring granularity while achieving high accuracy. Additionally, it can be used to predict future power usage by mining spatiotemporal patterns with minimal modifications to dataset construction building on the existing PowerDis framework. We evaluate PowerDis on both ARM-based and X86-based platforms. Extensive results show that PowerDis has great accuracy for power estimation and forecasting, reducing the mean absolute percentage error (MAPE) by up to 20% compared to other power modeling methods.