I decided to try ZFS-FUSE is my ongoing attempt to employ filesystem snapshots in my servers. ZFS-FUSE is in beta but I thought I would give it a try and see what happened. The install was not too difficult other than having to install a bunch of library dependencies along the way. These steps may work on other versions of Ubuntu, Debian, or other Debian derived distros but they have only been testing on Ubuntu 8.04.

I've concluded testing NILFS and found that even the latest version (2.0.5) is not ready for use. I was hoping to use NILFS to provide snapshots of virtual disks but I ran into nothing but trouble, I think due to the use of O_DIRECT by Xen which is not correctly supported by NILFS. My virtual machines with files on a NILFS filesystem complained of IO errors on their virtual disks. I like the design and tools of NILFS a lot so I will definitely keep an eye on it but it's not ready yet. Next on my list of new filesystems to try: ZFS-FUSE.

I was remarking to a colleague the other day that the reason there is so much filesystem development these days is that it is deeply understood, if not often expressed, that there are features which people want in a filesystem that are not available in any of the current filesystems. The problem, in my opinion, is that development efforts are so splintered that the lack of certain features does not lead to 1 or 2 new filesystems but far more. For disk file systems there is already Btrfs, NILFS, ext4, ChunkFS, Tux3, and more. And then there are cluster file systems…

All this reading about filesystems reminded me of a set of articles I read comparing a lot of the new Linux filesystems. The third part is really the most important except for details about how to actually get some of these running:

• Linux filesystems part 3: Comparing the standard filesystems

This roundup covered Ext2, Ext3, XFS, JFS, and ReiserFS for a baseline and then examined ChunkFS, NILFS, btrfs, Ext4, Reiser4, ZFS/FUSE, and ZFS/OpenSolaris. What I think most people don’t realize is that there have been a number of performance improvements in Ext3 in the 2.6 Linux kernel. In the real world tests Ext3 had the best performance of the total time for all tests. Of course, some of the numbers don’t look entirely right to me but the point is that Ext3 does pretty well, and is much faster than it used to be. Through in the fact that it is the most tested and compatible filesystem I think it’s an easy choice for systems. The features that would be nice that Ext3 is lacking are data checksumming and filesystem-level snapshots and, thus, I agree with the author that btrfs is one of the filesystems to watch.

In thinking about using ZFS under either Fuse/Linux or Nexenta/Solaris I think it's important to note why ZFS is of interest compared to the numerous other upcoming filesystems. I think, for me, what is attractive about ZFS are certain aspects of functionality such as checksumming and filesystem-based snapshots as well as the common set of tools for disk maintenance. While MD and LVM are great pieces of software, their tools are not always the best in my opinion. Although mdadm may be a step up from its predecessors it's not what I would consider intuitive. However, this is all theoretical as I have not actually used ZFS yet. And the servers I build and maintain will continue to be MD/LVM/Ext3 for some time I'm sure.

After trying my best to run xfs_repair on a large xfs partition (and failing) I realized that the decision to use xfs was a mistake. In no way should anyone use xfs unless: 1) they are sure there will be no compatibility issues, 2) they use it on a 64-bit system, 3) they use it on a system with lots of RAM, 4) they don't use it on a large partition, 5) they are not using it with block device encryption. Believe me, I've tried with each of these issues individually and in combination and have concluded that it was not worth the extra effort.