A Methodology for File System Performance Evaluation

This study presents a specification for the design of file system performance metrics. Standard system performance evaluation techniques involve the use of checkpoints imposed at the application layer. These checkpoints are part of a synthetic mix of programs, usually part of a benchmark, used to exercise a system. These evaluations compile system data, but in most cases only collect general information on the I/O subsystem. An important aspect of the I/O subsystem is a file system. In most performance studies the file system is measured based on system calls such as read, write, and delete. However, a file system is composed of many algorithms that perform the file system management duties. This study isolates those functions that compose a file system. It is important to provide a testing environment that isolates the subsystem used in the study. This study provides a controlled environment for data collection with the use of an operating system that allows the coexistence of multiple file systems. Based on this feature, a virtual model was developed as a specification for the design of checkpoints used to collect the appropriate raw data. This \u27data was then studied and analyzed. This study presents an implementation under the Linux Virtual File System model based on the virtual model design. This implementation was used to collect and present data on performance of several of the Linux file system implementations. This study shows that a controlled specification designed into the multiple file system support layer of an operating system provides: (J) a file system independent checkpoint that cannot be altered by the file system; (2) more file system related data by collecting information on most file system functions; and (3) a means for measuring file systems under live loads as well as synthetic loads