Abstract
Existing power modelling research focuses not on the method used for developing models but rather on the model itself. This paper aims to develop a method for deploying power models on emerging processors that will be used, for example, in cloud-edge data centers. Our research first develops a hardware counter selection method that appropriately selects counters most correlated to power on ARM and Intel processors. Then, we propose a two stage power model that works across multiple architectures. The key results are: (i) the automated hardware performance counter selection method achieves comparable selection to the manual selection methods reported in literature, and (ii) the two stage power model can predict dynamic power more accurately on both ARM and Intel processors when compared to classic power models.
Introduction
As computing shifts toward a "Cloud-Edge" model, data centers are becoming increasingly heterogeneous, mixing high-performance Intel Xeon processors with low-power ARM chips. Managing these systems efficiently requires fine-grained power monitoring. However, physical power meters often only provide "whole-system" data, leaving a gap in understanding how individual cores or specific software tasks consume energy. To fill this gap, researchers use Hardware Performance Counters (HPCs)—internal registers that track events like cache misses or instructions executed—to estimate power.
Methodology
Hardware Counter Selection (HCS) Method
The challenge with HPCs is twofold: there are dozens of counters, but hardware limits often allow only a few to be monitored at once. Choosing the "right" ones usually requires manual, architecture-specific expertise. Instead of relying on trial-and-error, we used a Random Forest (RF) algorithm to automate the selection of the top six most influential counters. By partitioning the dataset and averaging these importance values, we identified a set of counters that correlate most strongly with power consumption for a specific architecture.
Two-Stage Power Model (TSPM)
Simple models like Linear Regression (LR) are robust but often inaccurate, while complex models like Neural Networks (NN) or Support Vector Machines (SVM) frequently overfit, performing well on known tasks but failing on new, "unknown" workloads. To solve this trade-off, we built a hybrid Two-Stage Power Model:
- Stage 1: Uses Linear Regression to predict a "basic" power value. LR has low variance, meaning it doesn't fluctuate wildly when it sees new data.
- Stage 2: Uses an SVM to predict the residual (the difference) between the LR estimate and the actual measured power.
The final output is the sum of these two stages. This allows the model to capture non-linear complexities via SVM while staying anchored by the stability of the linear model.
Experimental Setup
The study utilized two distinct hardware platforms: an Intel Xeon Sandy Bridge server (Cloud Node) and an ODROID-XU+E board featuring the ARM big.LITTLE architecture (Edge Node). For workloads, we employed 16 scientific benchmarks, including Monte-Carlo simulations and Fast Fourier Transforms, providing a diverse range of computational intensities to test the models' robustness.
Results and Discussion
The Overfitting Trap
Our results confirmed that while sophisticated models like NN and SVM have lower error on "Known" data, they perform significantly worse than simple Linear Regression on "Unknown" workloads. This divergence is particularly visible on ARM architectures, where the complexity of the big.LITTLE design can lead to high variance in non-linear models.
Model Performance
The automated HCS method selected counters nearly identical to those chosen by human experts in previous literature, proving manual tuning is no longer necessary. Furthermore, the TSPM consistently outperformed classic models. On ARM, nearly 60% of "Unknown" workloads were predicted with less than 10% error using TSPM. On Intel, the TSPM matched the high accuracy of SVM for known tasks while maintaining the robustness of Linear Regression for unknown tasks.
Conclusion
This research demonstrates that automated counter selection and a two-stage hybrid modelling approach can provide accurate, stable power estimation across heterogeneous architectures. By modeling the residual of a linear prediction, we achieve a balance between precision and generalization essential for modern cloud-edge environments.
Summary updated