Análise das métricas SMART em HDDs para modelos de predição de falha nos discos usados em data centers para a nuvem

Abstract

Cloud data storage is one of the main activities of large cloud infrastructure providers. Objects (unstructured data) make up a large part of what is stored, and Hard Drives (HDs) are still the most widely used media to achieve significant amounts of storage at the best cost-benefit. The entire cloud service infrastructure must be redundant so that the data and the mechanism as a whole are reliable, durable, and available to the maximum. Thus, a significant aspect that has not yet been sufficiently explored and which is the subject of this work is the predictive maintenance of hard drives. In this work, the mechanisms behind this subject are explained by comparing redundancy strategies (Replication and Erasure Coding), monitoring SMART metrics and system reliability, calculating the Lifetime Value (RUL) of devices, and developing Machine Learning models to predict imminent failures in disks of a model. Furthermore, an experiment described in the article “Interpretable predictive maintenance for hard drives” [1] is replicated for a different disk model and range. Finally, considerations are gathered about the effectiveness of this model to treat the data in question and provide greater durability and reliability for a storage system based on hard drives. As a result, we seek to provide a broad knowledge of the subjects involved in this work.