IBM Israel Research Seminars

In this talk I will describe NFSv4.r, a replication extension to NFSv4 that is tuned for Grid workloads. Our measurements of Grid applications and benchmarks show that performance overhead due to NFSv4.r replication is negligible, even for workloads that are write-mostly, and even when most of the replication servers are distant.

NFSv4.r performance is achieved through hierarchical replication control, which takes advantage of spatial and temporal locality of reference, and by careful compromises on durability guarantees. However, consistency is not compromised: NFSv4.r guarantees close-to-open sharing semantics.

To understand the durability/performance tradeoff, we developed a calculus that takes into consideration the I/O characteristics of applications and the failure behavior -- unit failure and correlated failure -- of distributed storage nodes. With this, we are able to prescribe a file system replication strategy that maximizes the utilization of computational resources.

This work is joint research with Jiaying Zhang, an advanced doctoral candidate working under my supervision.

About the Speaker
Peter Honeyman is Research Professor of Information at the University of Michigan where he is Scientific Director of the Center for Information Technology Integration. As an experimental computer scientist, Honeyman builds middleware for file systems, security, and mobile computing. He has been instrumental in software projects including Honey DanBer UUCP, PathAlias, MacNFS, Disconnected AFS, and WebCard, the first Internet smart card. Current work centers on CITI's open source reference implementation of NFSv4 and its extensions for high end computing.