[Paper Review] Power Consumption Modeling and Prediction in a Hybrid CPU-GPU-MIC Supercomputer (preliminary version)

Alina Sîrbu, Özalp Babaoğlu|arXiv (Cornell University)|Jan 22, 2016

Parallel Computing and Optimization Techniques12 references3 citations

TL;DR

This paper proposes a data-driven power consumption modeling framework for hybrid CPU-GPU-MIC supercomputers using historical job traces and Support Vector Regression (SVR). It achieves over 80% prediction accuracy for user-specific job power consumption by leveraging job-level features such as application name, resource usage, and job duration, enabling real-time power-aware scheduling and billing without requiring low-level system instrumentation or code analysis.

ABSTRACT

Power consumption is a major obstacle for High Performance Computing (HPC) systems in their quest towards the holy grail of ExaFLOP performance. Significant advances in power efficiency have to be made before this goal can be attained and accurate modeling is an essential step towards power efficiency by optimizing system operating parameters to match dynamic energy needs. In this paper we present a study of power consumption by jobs in Eurora, a hybrid CPU-GPU-MIC system installed at the largest Italian data center. Using data from a dedicated monitoring framework, we build a data-driven model of power consumption for each user in the system and use it to predict the power requirements of future jobs. We are able to achieve good prediction results for over 80% of the users in the system. For the remaining users, we identify possible reasons why prediction performance is not as good. Possible applications for our predictive modeling results include scheduling optimization, power-aware billing and system-scale power modeling. All the scripts used for the study have been made available on GitHub.

Motivation & Objective

• To address the critical challenge of energy efficiency in ExaFLOP-scale HPC systems by enabling accurate job power prediction.
• To develop a user-centric, data-driven power modeling approach that does not require application code or low-level system monitoring.
• To support power-aware scheduling, billing, and system-level power modeling through predictive analytics.
• To identify and validate key job-level features influencing power consumption in hybrid CPU-GPU-MIC architectures.
• To enable real-time deployment of power prediction models with periodic retraining for adaptability.

Proposed method

• The study uses historical job traces from the Eurora supercomputer, collected at 5-second intervals, to extract job-level features such as application name, number of CPUs/GPUs/MICs used, and job duration.
• Power consumption is modeled as a regression problem using Support Vector Regression (SVR), with separate models trained per user based on their historical job data.
• Component-level power (CPU, GPU, MIC) is computed by summing power usage across nodes at 5-minute intervals, and total job power is the sum of component powers.
• The framework is trained using scikit-learn and validated using Google BigQuery for initial data analysis, with all scripts open-sourced on GitHub.
• A multiple-SVR model is used per user to capture individual usage patterns, improving over a baseline enhanced average model (EAM) that only considers component counts.
• The system supports real-time prediction by training models offline and updating them monthly, with low computational overhead due to per-user parallelization.

Experimental results

Research questions

• RQ1 Can job power consumption in a hybrid CPU-GPU-MIC supercomputer be accurately predicted using only high-level job features and historical data?
• RQ2 Which job-level features—beyond component count—most significantly influence power consumption in heterogeneous HPC systems?
• RQ3 How does user-specific modeling compare to application-level or average-based modeling in terms of prediction accuracy?
• RQ4 To what extent does data quality and variability affect the performance of power prediction models?
• RQ5 Can the proposed framework be deployed in real time for power-aware scheduling and system-level power modeling?

Key findings

• The proposed SVR-based user-specific model achieves prediction error under 20% for 80% of users, significantly outperforming the enhanced average model (EAM).
• The inclusion of application name, job duration, and concurrent resource usage by other jobs as features improves prediction accuracy beyond simple component-count models.
• Model performance degrades for 20% of users primarily due to data noise and inconsistent job naming or resource usage patterns.
• The framework is computationally efficient: global model training for all users takes only 2.92 core-hours, and inference on a full month of data takes just 6.81 minutes across all users.
• The method is scalable and parallelizable, with model retraining requiring only ~3 hours per month on a multi-core cluster, making it suitable for large-scale HPC systems.
• The approach enables practical applications such as power-aware scheduling, system-level power modeling, and user-side power estimation for billing and resource planning.