Re: finobt option for end user

[Alphazo <alphazo@xxxxxxxxx>]

Thanks for the detailed answer.

Regarding your last comment on mounting my volume read-only it might not be possible since I update the metadatas of my pictures quite often and thus even on older pictures. Part of my backup strategy, besides my unRAID NAS, that BTW has recently officially switched from ReiserFS to XFS, is a dedicated drive storing backup created by "bup" with the optional PAR2 checksum activated so I get incrementa and de-duplicated backup as well as protection against bit-rot.

PS: I was so excited about the new XFS metadata checksum that I went ahead and added a section about it to the ArchLinux official wiki. Now I'm trying to figure out if the finobt bit should be added as well under the performance section.

https://wiki.archlinux.org/index.php/XFS#Integrity

On Sun, Dec 21, 2014 at 9:38 PM, Michael L. Semon <mlsemon35@xxxxxxxxx> wrote:

On 12/20/14 17:52, Alphazo wrote:
> Hello,
>
> I'm pretty new to XFS. I'm considering moving away from ext4 to XFS because
> of the new self-describing option, performance and reliability improvements
> that XFS went through over the past year. Now I'm puzzled with the new free
> inode btree option (finobt). I tried to find some documentation about it
> but couldn't find the pros or cons. So from an end-user perspective with a
> couple of TB worth of photos:
> - Does it improve overall reliability?
> - Does it provide faster fsck/repair?
> - Does it improve any read or write operation?
> - Is it safe to use and does it recover as well as with finobt=0?
> - What is the typical case for enabling it and would you recommend using it
> for any new fs creation?
>
> Thank you in advance for your pointers.
> Alphazo

As a user...

You could try it out and see:

mkfs.xfs -m crc=1,finobt=1 <block_device_to_format>

Here, on old junk hardware, finobt aids greatly in multitasking on
small-file creation. It may not have the same effect on your hardware.

finobt has given very, very little trouble since it was accepted into the
kernel code and xfsprogs.

Recovery and reliability have become roughly equal between old XFS, new
XFS, and new XFS with finobt. They just approach the problem in different
ways. With old XFS, I rely heavily on xfs_repair and xfsdump to show
filesystem issues. With new XFS, the kernel code is more likely to complain
as the issues are happening.

finobt introduces the slightest bit of overhead that might be noticed at the
point of resource exhaustion. Otherwise, its overhead might be hard to spot.

As for safety and reliability, I'm working with kernel 3.18 with the drives'
write caches shut off and using the latest xfsprogs. It works great. The
new v5-superblock XFS progresses at a fast pace, and it helps greatly to
keep up with it. kernel 3.18 seems to be a better kernel, period. YMMV.

Try a test filesystem first and test to your liking first. If something
goes wrong, you should find that out in two weeks of testing the file
system.

Just curious, though, with that volume of images, is there a possibility of
making a read-only filesystem? That would take a lot of the safety and
recovery issues out of the equation, even if you stick with ext4. It would
make your main concern to sha256sum the files so that you can check them
later for bit-rot.

Good luck!

Michael