Most traditional cluster file systems replicate whole files to support the data availability. However, considering that the access patterns of file systems have high localities, replicating whole files is not an efficient strategy to improve the data availability. This paper presents the design and the implementation of HA-PVFS, a PVFS supporting high data availability, in which the replication is performed partially based on the locality information. Our performance results show that HA-PVFS is dynamic with little overhead over various access patterns.
Published in:
Computer and Information Technology, 2007. CIT 2007. 7th IEEE International Conference on
Date of Conference: 16-19 Oct. 2007