NP: An N-Dimensional Feature Space and P-Partitioning Subdomain Processor Power Modeling
Juan Chen , Aolin Cao , Danchen Huang , Shiwen Liang , Zhaoyang Ma , Rongyu Deng , Shuohao Wang , Song Gao
With the growing energy consumption challenges in modern computing systems, developing high-precision processor power models has become a critical enabler for system-level power optimization and management. However, existing approaches exhibit significant limitations in feature selection, heterogeneous dataset processing, and model adaptability—each of which hinders progress in modeling accuracy. This paper proposes NP-Mdl, a processor power modeling approach that partitions the N-dimensional feature space into P subregions, along with its corresponding generation algorithm NP-Alg. The method employs a two-stage feature screening mechanism that integrates mutual information and Random Forest to mitigate noise, uses density peak clustering to partition highly heterogeneous sample sets, and adaptively selects optimal predictive models for each subregion. Experiments were conducted on x86-based and ARM-based platforms using benchmarks including SPEC2017, HPL-AI, and HPCG. The results demonstrate that NP-Mdl achieves Mean Absolute Percentage Errors (MAPE) of 2.73% and 2.65% on the two platforms, representing a significant reduction from the baseline MAPE of 6.98% and 6.88%, respectively. When applied to dynamic power management, the NP-Alg algorithm effectively constrains processor power below the predefined threshold, incurring a performance loss of only 0.98%. while achieving a significant 6.2% reduction in energy consumption. These results further validate the superior effectiveness and practical applicability of the NP-Alg approach in real-world scenarios.
Processor Power Modeling / Feature Selection / Region Partitioning / Adaptive Model Selection
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(21379KB)
