[Paper Review] Computing server power modeling in a data center: survey,taxonomy and performance evaluation

Leila Ismail, Huned Materwala|arXiv (Cornell University)|Apr 26, 2020

Cloud Computing and Resource Management89 references6 citations

TL;DR

This paper presents a unified evaluation of 24 software-based server power models across three server architectures, using consistent benchmarks, measurement techniques, and error metrics. It finds that SVM-based models outperform others in multi-variable scenarios, while interpolation is best for single-variable models, with performance varying significantly across hardware platforms.

ABSTRACT

Data centers are large scale, energy-hungry infrastructure serving the increasing computational demands as the world is becoming more connected in smart cities. The emergence of advanced technologies such as cloud-based services, internet of things (IoT) and big data analytics has augmented the growth of global data centers, leading to high energy consumption. This upsurge in energy consumption of the data centers not only incurs the issue of surging high cost (operational and maintenance) but also has an adverse effect on the environment. Dynamic power management in a data center environment requires the cognizance of the correlation between the system and hardware level performance counters and the power consumption. Power consumption modeling exhibits this correlation and is crucial in designing energy-efficient optimization strategies based on resource utilization. Several works in power modeling are proposed and used in the literature. However, these power models have been evaluated using different benchmarking applications, power measurement techniques and error calculation formula on different machines. In this work, we present a taxonomy and evaluation of 24 software-based power models using a unified environment, benchmarking applications, power measurement technique and error formula, with the aim of achieving an objective comparison. We use different servers architectures to assess the impact of heterogeneity on the models' comparison. The performance analysis of these models is elaborated in the paper.

Motivation & Objective

• To enable objective, reproducible comparison of software-based server power models by standardizing experimental conditions.
• To evaluate the impact of hardware heterogeneity on power model accuracy across different server architectures.
• To identify the most accurate power modeling techniques—both linear and non-linear—for diverse application workloads.
• To provide actionable insights for data center operators on selecting or developing energy-efficient power models.
• To highlight critical design considerations such as idle power, feature selection, and kernel function choice in power modeling.

Proposed method

• Conducted experiments on three distinct server architectures to assess model robustness across hardware heterogeneity.
• Used nine diverse benchmarking applications to simulate CPU-, memory-, disk-, and network-intensive workloads.
• Employed a consistent power measurement technique and error calculation formula (MAPE) across all models and systems.
• Evaluated 24 software-based power models, including linear regression, SVM, Lasso, and interpolation, using system performance metrics and counters.
• Applied variable selection techniques to identify high-correlation features (e.g., CPU, memory, disk I/O) for model development.
• Tested both throughput-based and performance counter-based modeling approaches, comparing their accuracy and calibration needs.

Experimental results

Research questions

• RQ1 How do different software-based power models perform under a unified experimental setup across heterogeneous server architectures?
• RQ2 Which modeling technique—linear or non-linear—yields the lowest estimation error for server power consumption?
• RQ3 How does the inclusion of memory utilization and multiple performance counters affect model accuracy compared to CPU-only models?
• RQ4 What is the impact of server idle power assumptions on model prediction accuracy, especially for energy-efficient systems?
• RQ5 How do kernel function selection and polynomial utilization functions (linear, quadratic, cubic) influence the performance of SVM-based models?

Key findings

• Among single-variable models, interpolation achieved the lowest estimation error, outperforming linear regression and other methods.
• Among multi-variable models, the SVM-based power model demonstrated the lowest error, particularly for complex, multi-resource workloads.
• For CPU-intensive applications, the interpolation model provided the most accurate predictions, while SVM excelled for CPU+memory, CPU+memory+disk, and CPU+memory+disk+network workloads.
• The Lasso regression model with 30 variables performed the worst, exhibiting the highest estimation error across all application types.
• Model performance varied significantly across different server architectures, indicating that model portability across platforms is limited without re-evaluation.
• The assumption that idle power is 70% of peak power leads to misleading predictions, especially for modern, energy-efficient servers with lower idle power ratios.