Section: New Results

Towards a BLOB-based file system

Participants : Viet-Trung Tran, Gabriel Antoniu, Bogdan Nicolae.

Most object-based file systems exhibit a decoupled architecture that generally consists of two layers: a low-level object management service, and a high-level file system metadata management. We have explored how this two-layer approach could be used in order to build an object-based grid file system for applications that need to manipulate huge data, distributed and concurrently accessed at a very large scale. We have investigated this approach by experimenting how the Gfarm grid file system developed at the University of Tsukuba could leverage the properties of the BlobSeer distributed object management service, specifically designed for huge data management under heavy concurrency.

We thus leverage Gfarm's powerful file metadata capabilities and rely on BlobSeer for efficient and transparent low-level distributed object storage. The goal is to build a BLOB-based grid file system that exhibits scalable file access performance in scenarios where huge files are subject to massive, concurrent, fine-grain accesses. Instead of using the local disk for data storage, each Gfarm storage node stores data in BlobSeer. The benefits are mutual: by delegating object management to BlobSeer, Gfarm can expose efficient fine-grain access to huge files and benefit from transparent file striping (TB size). On the other hand, BlobSeer benefits from the file system interface on top of its current API. We have defined and implemented a preliminary integrated architecture, and we have evaluated it through a series of preliminary experiments conducted on the Grid'5000 testbed. The resulting BLOB-based grid file system exhibits scalable file access performance in scenarios where huge files are subject to massive, concurrent, fine-grain accesses. These results have been published in [10] .