Application of multivariate time-series model for high performance computing (HPC) fault prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• High-Performance Computing

Background:

• Supercomputing systems face increasing reliability demands due to their scale and complexity.
• Effective fault prediction is crucial for maintaining the operational integrity of these systems.
• Existing methods struggle to capture complex spatio-temporal fault patterns.

Purpose of the Study:

• To propose a multidimensional fusion fault prediction model for large-scale supercomputing systems.
• To enhance the accuracy and efficiency of fault prediction in complex computing environments.
• To address the need for high reliability in advanced computational infrastructure.

Main Methods:

• Developed a novel CBA-net (CNN-BiLSTAM-Attention) model incorporating HDBSCAN clustering for data preprocessing.
• Utilized deep learning to extract and learn spatial and temporal features from fault logs.
• Focused on sensitivity to time-series features and local feature extraction.

Main Results:

• Achieved a Root Mean Square Error (RMSE) of 0.031 for fault occurrence time prediction.
• Attained an average prediction accuracy of 93% for fault occurrence node location.
• Demonstrated fast convergence and improved fine-grained fault prediction capabilities.

Conclusions:

• The proposed CBA-net model effectively predicts faults in large supercomputers.
• The model's ability to extract spatio-temporal features leads to high prediction accuracy.
• This approach significantly enhances the reliability and fine-grained fault prediction for supercomputing systems.