DHISC : Disk Health Indexing System for Centers of Data Management

Abstract

If we want to have reliable data centers, we must improve reliability at the lowest level of data storage at the disk level. To improve reliability, we need to convert storage systems from reactive mechanisms that handle disk failures to a proactive mechanism that predict and address failures. Because the definition of disk failure is specific to a customer rather than defined by a standard, we developed a relative disk health metric and proposed a customer-oriented disk-maintenance pipeline. We designed a program that processes data collected from data center disks into a format that is easy to analyze using machine learning. Then, we used a neural network to recognize disks that show signs of oncoming failure with 95.4-98.7% accuracy, and used the result of the network to produce a rank of most and least reliable disks at the data center, enabling customers to perform bulk disk maintenance, decreasing system downtime.