Assessing Predictive Models for Energy Consumption Across Varied Software Environments

Problem and motivation

Energy models are sensitive to the selected counters, operating-system behavior, hardware, and workload. A model that performs well within one application can therefore be misleading when used on a different software environment or on unseen data.

Method and contribution

The study measures energy with Intel RAPL and collects performance counters with perf on a Ryzen 7 5825U system running Ubuntu 22.04.2 at the stated 1.6 GHz configuration. Stress-ng supplies controlled CPU workloads with 32 tasks and 6,528 combinations; Firefox versions 91, 96, 101, 105, 106, and 111 and MongoDB 7.0.0–7.0.4 provide application cases. The evaluated regressors include linear regression, LASSO, Ridge, SGD, MLP, and random forest, with repeated measurements, MAPE evaluation, and comparisons between event-based and alternative feature representations.

Findings and evidence

The results show no universal best model: within-application performance is often strongest for hybrid feature sets, while event-based representations are more promising for transfer to unseen software conditions. The reported application-level MAPE values include 7.2%, 4.51%, and 6.58% for Firefox cases and 2.91% for the MongoDB case, but the error can exceed 100% for some model/feature combinations. The central empirical result is therefore the dependence of prediction quality on the application and feature representation, not a single universally transferable predictor.

Limitations and future directions

Limitations: The measurements use one laptop-class CPU, one operating-system configuration, Intel RAPL, selected software versions, and controlled workloads. Cross-application splits, counter availability, frequency behavior, and model hyperparameters constrain external validity; software energy outside the tested inference/workload conditions is not established.

Future work: Replicate the comparison on heterogeneous CPUs, GPUs, operating systems, frequency policies, and larger application suites; report confidence intervals and leakage-safe cross-application splits; and evaluate online calibration under workload drift and concurrent activity.

Resources

• Code

Sources and identifiers

• Published version published
• IEEE record public_source_record

When to cite this paper

Cite this paper when evaluating whether software-energy models transfer across applications or feature representations.

• Perf/RAPL-calibrated regression and neural energy predictors.
• Event-based versus alternative features for unseen-software generalization.
• The Firefox/MongoDB and Stress-ng evaluation design.
• The finding that no single model is universally best across software environments.

Citation

@inproceedings{ezzatiJivan2024assessingpredictive,
  author = {Tong Zhang and Sarwat Islam Dipanzan and Leila Tahmooresnejad and Naser Ezzati-Jivan},
  title = {Assessing Predictive Models for Energy Consumption Across Varied Software Environments},
  year = {2024},
  booktitle = {2024 IEEE International Conference on Big Data (BigData)},
  pages = {5233-5242},
  publisher = {IEEE},
  doi = {10.1109/BigData62323.2024.10825500},
  url = {https://doi.org/10.1109/BigData62323.2024.10825500}
}

Zhang, T., Dipanzan, S. I., Tahmooresnejad, L., & Ezzati-Jivan, N. (2024). Assessing Predictive Models for Energy Consumption Across Varied Software Environments. In 2024 IEEE International Conference on Big Data (BigData) (pp. 5233-5242). https://doi.org/10.1109/BigData62323.2024.10825500

T. Zhang, S. I. Dipanzan, L. Tahmooresnejad, and N. Ezzati-Jivan, "Assessing Predictive Models for Energy Consumption Across Varied Software Environments," in 2024 IEEE International Conference on Big Data (BigData), pp. 5233-5242, 2024, doi: 10.1109/BigData62323.2024.10825500