DRAM Failure Prediction with Multiple Feature Fusion and Screening Mechanism

In the era of prevalent cloud computing, cloud servers have been widely utilised. Dynamic Random Access Memory (DRAM) failure is one of the main causes of server downtime. In this study, a multi-feature fusion and screening mechanism for DRAM failure prediction is proposed. Firstly, the Correctable Errors (CEs) logs were reconstructed into spatial feature maps and the ResNet-Transformer network was constructed for deep feature extraction. Then novel manual features were extracted and the hardware information of the server was quantified. After fusing all the features, a rigorous feature selection was performed and finally DRAM failure prediction was performed. We explored the effects of different classifiers, sample scales and prediction windows. This study obtained good results on Large-scale Dataset collected in the real cloud centre.