Machine Learning Based Collaborative Prediction of SSD Failures in the Cloud

SSDs (Solid-State Drives) have become integral components in modern data centers. Under such massive deployment, ensuring their reliability, longevity, and optimal performance is crucial. Despite SSD technology and architecture advancements, accurately predicting their failures, particularly with imperfect real-world data, remains a pertinent research challenge. Dataset imbalances have led to suboptimal prediction accuracy in baseline models. This study introduces to incorporate model-wise class balancing, aiming to refine the data processing for improved accuracy in machine learning models for SSD failure detection. When tested on Alibaba’s dataset of 700k NVMe SSDs, this method yielded higher failure prediction accuracy, with the average recall increasing from 51% to 63% and precision scores rising from 59% to 78%. This improvement in recall and precision demonstrates the method’s potential to advance the field of SSD failure prediction.