Disk Failure Analysis And Prediction Based On Disk I/O Load

In the era of information, there is much new information generating every day. Thetotal amount of the global data is increasing at a rate of50%per year. In the rate of growth,the total amount will increase to1000-fold every twenty years. More and more data isstored in the data center. Data center provide a large storage space through theaccumulation of hard disks, and these disks require lots of servers to host. Any serverfailure will affect the availability of storage space. According to statistics, the hard drive isthe main failure component of the server, so it is very important to improve the availabilityof hard disk without affecting the data access. While the troubleshooting and othertraditional means cannot solve such problems.The disk failure analysis and prediction system based on disk I/O load is designed tosolve these problems as mentioned before. It throws light upon the change rule betweendifferent server configurations and hard disk failure rate in its life cycle through thestatistics and cluster analysis on its several factors associated closely with disk failure, andprovides the fundamental basis for the development of maintenance programs. With thedisk I/O load and the disk failures collected, the special characteristics of hard disk beforeand after failure is analyzed, the disk I/O load characteristics of the failure classes andregular classes are trained, the two types of samples are mapped into a high dimensionalfeature space by the use of Support Vector Machine which is a machine learning method,the two classes with a boundary are divided and thus a failure prediction model can begenerated. With real-time collection of all servers’ hard disk I/O load and the use of failureprediction model, the system predicts the disk failure dynamically, thus avoiding theoccurrence of disk failures and ensuring the normal operation of the system.The experimental results show that there is a strong correlation between the diskfailure rate and the server configuration information. It is possible to create an effectivedisk failure prediction model depending on the disk I/O load. Besides, the performancetests show that the disk failure prediction with statistical characteristics of disk I/O in thisway can reach a Detection Rate up to73.7%while keeping a tolerable False Alarm Ratewhich is controlled below3%.