Job Power Profile Prediction and Analysis (2021 - 2023)

Overview

This project classified HPC jobs based on time-series power consumption to enhance descriptive analytics and train predictive models for data center cooling.

• Developed a machine learning pipeline to classify job power profiles based on consumption patterns.
• Designed a feature extraction framework that distills 186 features from time-series power data.
• Built an open-set neural network classifier to categorize jobs, identifying known and unseen workload patterns.
• Applied generative adversarial networks (GANs) to improve clustering and refine classification models.
• Provided system-wide insights into HPC power usage, informing energy-aware resource management.

Technology

• GAN, k-means, Open-set classification
• Scikit-learn, PyTorch
• Apache Airflow
• OLCF Summit Supercomputer Telemetry

Publication

• Karimi et al., “Power Profile Monitoring and Tracking Evolution of System-Wide HPC Workloads”, ICDCS'24, 2024