Research On Predict Disk Failure Based On S.M.A.R.T.

In recent years, cloud computing technology is developing rapidly, which brings changes to the business model for many companies, but also to the ways of working for individuals. However, cloud security incidents continue to occur, often with serious consequences of data loss, bring serious losses to companies and individuals, but also hindered the widely use of cloud computing. The security and reliability of cloud computing has become the focus of attention of companies and individuals.To ensure that data is not lost, we must focus on the cloud storage security first. Cloud storage is the foundation of cloud computing, and the core of cloud storage is disk drive. Because of the progress of disk drive’s manufacturing technology, the probability of disk failure becomes lower and lower. However, due to the number of disk drives in the cloud storage is extremely large, disk failure still occur frequently in a cloud environment. Because disk failure occurs frequently, so our data is at risk. And because the disk failure rate is very low, so the work of predicting disk failure and maintaining cloud platform of operation and maintenance personnel becomes extremely difficult.Self-Monitoring, Analysis and Reporting Technology (S.M.A.R.T.) is one of the standard conditions stipulated by ATA standard and all disk manufacturers must be followed. It can determine the health status of the disk by monitoring the motor, head and temperature and so on and compare them with a security threshold which set by disk manufacturer when disk is running. When determine a fault occurs, it can automatically warn the user, some even can simple repair automatically, such as audo deceleration and backup data. At precent, threshold determination method which based on S.M.A.R.T. is the disk failure predition method widely used by disk manufaturers. However, the true positive rate (TPR) of this method is only 3-10%, the value is too low, so, the effectiveness of warning is not practical. Now, research on predict disk failure besed on S.M.A.R.T. is few, the models have been established are besed on disk manufacturer’s S.M.A.R.T. data and other data like environment, and are difficult to be applied to predict disk failure in a actual user cluster.In this paper, we built an effective disk failure predication model using S.M.A.R.T. alone based on actual data observated by user. The main work is summarized as follows:(1) Study the technology of disk S.M.A.R.T., and analysis the method of predict disk failure based on S.M.A.R.T. data.(2) Collect and preprocess disk S.M.A.R.T. data from actual cloud computing platform.(3) Classified disk status as "normal" and "will fail within 24 hours" using disk repair records. Disk failure is definded as that:a disk drive is considered to have failed if it was recognized the need for replaced as part of a repairs procedure. And the recognized time is the time when filure occurs.(4) For the actual disk failure rate is extremely low, and the experimental objects are extremely unbalanced datasets, we propose a hybrid disk failure predication method called DKSS. This method is a combination of clustering, resampling and integrated classifiers based on actual disk S.M.A.R.T. data of user cluster. In the stage of train model, we reconstruct the dataset by clustering and resamping to balance the positive samples and negative samples. And then use support vector machine method to learn and train some child classifiers. At last, poll and integrate the results of all child classifiers. In the stage of test model, use clustering method to reduce the size of sample first, and then use the integrated classifier to predict.(5) Use DBSCAN, K-means, SMOTE and SVM to build the DKSS model, Prove the validity of the DKSS model with experiment and analysis the predictive performance of this model.The innovation of this paper is:(1) Based on actual user environment, using only disk S.M.A.R.T. data to predict disk failure.(2) Porposed DKSS mixed strategy, and applied to predict disk failure. Experiments show that, this method has good prediction performance and strong generalization ability.