Computing Node Failure Prediction Based on Continuous-Time Dynamic Graph.

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Large model training and inference rely on multinode cooperation.
• Computing node failures cause significant overhead, data loss, and service interruptions.
• Current failure prediction models struggle with causal relationships and spatial-temporal correlations.

Purpose of the Study:

• To develop an accurate computing node failure prediction scheme.
• To address limitations of existing time-series models in capturing complex correlations.
• To mitigate overhead and service disruptions in dynamic cluster environments.

Main Methods:

• A continuous-time dynamic graphs-based computing node failures prediction (CTDG-NFP) scheme was designed.
• A novel multi-dimensional feature-biased neighbor sampling method was introduced.
• Diverse computing node failure motifs were extracted using feature-biased walks and anonymization, then encoded with a time encoder.
• Contrastive learning was employed to train the prediction model.

Main Results:

• The CTDG-NFP scheme demonstrated superior performance across six key metrics.
• Evaluations used real-world failure traces, confirming effectiveness.
• The method successfully captured complex spatial-temporal correlations and causal relationships.

Conclusions:

• The CTDG-NFP scheme offers a significant advancement in predicting computing node failures.
• This approach enhances the reliability and efficiency of large model training and inference.
• The proposed method provides a robust solution for dynamic cluster environments.