$AP^{3}$: Adaptive Power Prediction Framework based on Spatial Partition Multi-Phase Model
Juan Chen, Zhixin Ou, Yifei Guo, Xinxin Qi, Yuyang Sun, Lin Deng, Hongyu Chen, Zihan Lin · 2021
The accuracy of processor power modeling is an important foundation for power management and optimization on parallel computing system. It is difficult to build a high-accuracy instantaneous CPU/DRAM power prediction model. One of the main reasons for the low accuracy is that the processor architecture sometimes influence greatly on the accuracy of power prediction. For example, the accuracy of processor power model is affected by inaccurate modeling of the uncore part of processor. Another reason comes from the limitation of static models for instantaneous power prediction. Despite the existence of various optional linear/nonlinear models, the fixed training set and model coefficients are insufficient to make high-accuracy instantaneous predictions for various target programs. Aiming at the above two issues, this paper proposes an Adaptive Power Prediction framework based on spatial Partition multi-phase model ($AP^{3}$). Spatial partition mainly solves the impact of uncore power on the prediction accuracy, and adaptability solves the limitation of the static model on the power prediction accuracy. According to the experimental results on both ARM-based and x86-based processor platforms,$AP^{3}$greatly increases the CPU and DRAM power instantaneous prediction accuracy. Spatial partition reduce the prediction error (MRE) by 0.3%–8.2% compared to previous single model, while adaptive update further reduces the error (MRE) by 1.7%–7.1% compared to previous static model.