The Umbrella File System: Storage Management Across Heterogeneous Devices

Abstract

With the advent of Flash based solid state devices (SSDs), the differences in physical devices used to store data in computers are becoming more and more pronounced. Effectively mapping the differences in storage devices to the files, and applications using the devices, is the problem addressed in this dissertation. This dissertation presents the Umbrella File System (UmbrellaFS), a layered file system designed to effectively map file and device level differences, while maintaining a single coherent directory structure for users. Particular files are directed to appropriate underlying file systems by intercepting system calls connecting the Virtual File System (VFS) to the underlying file systems. Files are evaluated by a policy module that can examine both filenames and file metadata to make decisions about final placement. Files are transparently directed to and moved between appropriate file systems based on their characteristics. A prototype of UmbrellaFS is implemented as a loadable kernel module in the 2.4 and 2.6 Linux kernels. In addition to providing the ability to direct files to file systems, UmbrellaFS enables different decisions at other layers of the storage stack. In particular, alternate page cache writeback methods are presented through the use of UmbrellaFS. A multiple queue strategy based on file sequentiality and a sorting strategy are presented as alternatives to standard Linux cache writeback protocols. These strategies are implemented in a 2.6 Linux kernel and show improvements in a variety of benchmarks and tests.