Research on memory failure prediction based on ensemble learning.

Area of Science:

• Computer Science
• Data Center Reliability
• Machine Learning

Background:

• Predicting memory failures is critical for data center stability.
• Current single-classifier methods often yield inaccurate predictions.
• Correctable error (CE)-driven memory failures can cause server downtime.

Purpose of the Study:

• To develop an advanced ensemble model for predicting CE-driven memory failures.
• To improve the accuracy and stability of memory failure predictions.
• To enhance the overall reliability of data center operations.

Main Methods:

• Proposed a novel ensemble model combining multiple classifiers (Random Forest, LightGBM, XGBoost).
• Implemented a weighted approach for classifiers based on individual performance.
• Optimized the decision-making process within the ensemble model.

Main Results:

• Achieved over 84% accuracy in predicting memory failures.
• Outperformed existing single and dual-classifier models in validation tests.
• Demonstrated excellent predictive performance using real-world data center data.

Conclusions:

• The proposed ensemble model offers a more accurate and stable solution for predicting CE-driven memory failures.
• This approach enhances data center operational stability by enabling timely intervention.
• The findings confirm the model's effectiveness and potential for practical application in large-scale data centers.