Scholarlink

<p>With the ever-increasing scale of High-Performance Computing (HPC) systems, energy consumption has become a key challenge for sustainable computing. Accurate prediction of HPC job power consumption is essential for power-aware resource allocation and energy-efficient scheduling. This paper proposes a Feature Enhancement and Machine Learning (FEML) framework for HPC job power consumption prediction. Using only job submission information and historical logs, FEML integrates textual and numerical features through a feature enhancement module to obtain more expressive feature representations, and combines the enhanced and original features to improve prediction performance. Multiple machine learning regressors are evaluated for offline prediction, and a time-weighted online sliding-window mechanism is further introduced to adapt to temporal variations in dynamic environments. Offline experiments on four HPC datasets with different characteristics show that FEML tends to achieve stronger performance on datasets with richer textual submission metadata and larger numbers of job samples, while online experiments further demonstrate that the time-weighted online sliding-window mechanism can reduce prediction errors under evolving workloads. These results indicate that FEML is effective for HPC job power consumption prediction and provides a practical basis for future studies on power-aware scheduling and energy-efficient computing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feml: a feature enhancement and machine learning framework for HPC job power consumption prediction

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Heng Guo,
  • Xiaojun Bai,
  • Ji Zhou,
  • Wei Zhang,
  • Jianqiang Huang

摘要

With the ever-increasing scale of High-Performance Computing (HPC) systems, energy consumption has become a key challenge for sustainable computing. Accurate prediction of HPC job power consumption is essential for power-aware resource allocation and energy-efficient scheduling. This paper proposes a Feature Enhancement and Machine Learning (FEML) framework for HPC job power consumption prediction. Using only job submission information and historical logs, FEML integrates textual and numerical features through a feature enhancement module to obtain more expressive feature representations, and combines the enhanced and original features to improve prediction performance. Multiple machine learning regressors are evaluated for offline prediction, and a time-weighted online sliding-window mechanism is further introduced to adapt to temporal variations in dynamic environments. Offline experiments on four HPC datasets with different characteristics show that FEML tends to achieve stronger performance on datasets with richer textual submission metadata and larger numbers of job samples, while online experiments further demonstrate that the time-weighted online sliding-window mechanism can reduce prediction errors under evolving workloads. These results indicate that FEML is effective for HPC job power consumption prediction and provides a practical basis for future studies on power-aware scheduling and energy-efficient computing.