Xuesen Tian

Xuesen Tian

Auto-generated profile, based on publicly available information

Xuesen Tian

Qilu University of Technology (Shandong Academy of Sciences), China and Shandong Computer Science Center (National Supercomputing Center in Jinan), China

Research Focus & Metrics

Representative Publications

View All Publications
17 June 2023
Verified
Conference Paper
Open access
XTXuesen Tian
Xuesen Tian
XLXiang Li
Xiang Li
JZJie Zhang
Jie Zhang
ZZZhigang Zhao
Zhigang Zhao
CWChunxiao Wang
Chunxiao Wang
JWJibin Wang
Jibin Wang
This paper proposes a high-performance computing (HPC) job power consumption prediction framework based on online incremental learning. The framework combines traditional offline models with online incremental update mechanisms, allowing it to quickly adapt to the uncertainty and dynamic changes of power consumption behaviors in HPC jobs. We use NBeats, LSTM, Seq2seq, and GRU models as our offline models and combine them with the proposed online update mechanisms to achieve online incremental updates. We conducted experiments using real HPC job data, and the results show that our online incremental model can predict more accurately than traditional offline models. In addition, We propose an update frequency control mechanism that reduces the running time by about 30 while guaranteeing the accuracy of the prediction. This indicates that the HPC job power consumption prediction framework based on online incremental learning proposed in this paper has a wide range of applications in practical use.