Keywords

• Journalized filesystems
• Distributed filesystems
• Scalability, parallelism, availability
• Filesystem features in userland
• API change proposal for Linux 2.5.*

Motivation

After years of intensive usage, ext2 is a proven fast, and quite resilient filesystem. Yet, the exponential growth of available storage space and new uses (for example mid-range servers) revealed some weaknesses, especially:

• after unclean shutdowns, e2fsck can take an awfully long time to check large filesystems. When it's done, filesystem meta-data state is not totally forseeable;
• ext2 does not scale well to large directories;
• ext2 does not scale very well to large files;
• etc.

The first problem can be solved without giving up too much performance, using journalized filesystems. Other problems can be widely alleviated by switching to high performance on-disk data structures. Many native and foreign journalized filesystem projects are currently in the works.

These projects, in turn, revealed that current Linux block buffering features (page cache, buffer cache, LVM layer, software RAID layer,...) may need some rework to accomodate their needs in an unified way. The VFS interface also needs enhancements to bring new features to the upper layers (larger bitwidth, improved file access rights, optimized interfaces for database and multimedia applications, copy-on-write hard links,...), with improved parallelism.

High performance, high availability distributed filesystems are attractive to client-server, clustering or even mobile uses.

When implemented on top of a journalized filesystem, they may take advantage of the journals to restore the integrity of the distributed data. Could a standardized journal API allow to independently choose local and distributed filesystems?

Many filesystem features like filenames rewriting, compression, encryption, auditing, and other less conventional tricks can often be implemented in higher layers (VFS or userland), independently of the "storage" details.

Keeping good performance requires hooks at strategic points.

Sub-topics

• ext3, Reiserfs, IBM's JFS, SGI's XFS
• Coda, OBDFS/Intermezzo (Lustre project)
• Podfuk, stackable filesystems
• Block buffering
• Virtual File System

Program

Thursday, July 6th

Friday, July 7th

Mailing-list for topic organisation

• To subscribe, send an e-mail to sympa@lsm.abul.org with "subscribe lsm-5" within the message body.
• To post to the list, send e-mail to lsm-5@lsm.abul.org.