Predicting Hard Drive Failures for Cloud Storage Systems

Abstract

To improve reactive hard-drive fault-tolerance techniques, many statistical and machine learning methods have been proposed for failure prediction based on SMART attributes. However, disparate datasets and metrics have been used to experimentally evaluate these models, so a direct comparison between them cannot readily be made. In this paper, we provide an improvement to the Recurrent Neural Network model, which experimentally achieves a 98.06% migration rate and a 0.0% mismigration rate, outperforming the state-of-the-art Gradient-Boosted Regression Tree model, and achieves 100.0% failure detection rate at a 0.02% false alarm rate, outperforming the unmodified Recurrent Neural Network model in terms of prediction accuracy. We also experimentally compare five families of prediction models (nine models in total), and simulate the practical use.