The ability to efficiently find, open and update files is key to high performance computing systems. Consequently, the file metadata, which are the information needed to open a file and locate its data, must also be efficiently managed. In parallel file systems an additional level of complexity stems from the possible scattering of metadata and client applications over a network of cluster nodes, making the metadata management a critical part of the file system. A novel approach to metadata management in a parallel file system is implemented in AMPVFS. AMPVFS is a new experimental parallel file system that uses hardware implemented distributed shared memory for metadata storage and distribution. AMPVFS is based on two existing products: the open source parallel file system PVFS, and a gigabit network AmpNet. |