Preview

PDF, English Download (825kB) | Terms of use

Abstract

In today's file systems each file is made up of data and metadata. The metadata contains some information about the associated data, like ownership and permissions of the file. While this usually is useful, there are situations when the additional overhead of such a design becomes a problem in terms of performance. This is especially true for cluster file systems, because due to their design every metadata operation is even more expensive. If a user creates several thousand temporary files that are going to be deleted soon anyway, it is not necessary to store detailed information about them. In this thesis several changes are made to the parallel cluster file system PVFS to better deal with such cases. To do this, PVFS is altered such that certain unnecessary metadata is discarded and therefore metadata performance is increased. Several tests with a large quantity of files are done to measure the benefits of these changes. The reduction of the metadata overhead halves the time needed for some common file system operations. The speedup of those operations is also analyzed in detail by visualizing the internal workflow of PVFS. Also, additional work that could be done to further increase the metadata performance as well as possible additions to the actual implementation are presented.