Cloud Storage Behavior Analysis Using Time Series Clustering

To meet with different storage requirements, lots of products are provided by cloud service providers. These products vary greatly in price and performance. Choosing different storage services to build tiered storage systems could reduce overall costs of using of cloud services. But choosing of storage services is not easy because it must meet storage performance requirements and minimize the costs. We use machine leaning methods to solve this problem. Log files that keep storage access traits are used to analyze storage access patterns. Firstly, we process log files and generate access frequency time series, which are translated to N-Hot time series later. Then we extract feathers from N-Hot time series and use K-Means clustering method to classify storage objects. Different migration policies could be made to optimize storage usages according to these different classes of storage objects. The experiments show that our approach is useful, well performance and scalable.