Re: XFS worth it?

["Brian D. Winters" <brianw-debian@pteranodon.org>]

Just adding another data point...

On Fri, Apr 15, 2005 at 11:51:38PM +0200, Thomas Steffen wrote:
> * The log is not compatible between i386 and AMD64.

I have run XFS on a number of machines of both architectures for quite
a while (~two years i386, ~one year AMD64), but I can't say I've ever
tried to take a crashed volume from one to the other.  I generally
recover it on the same computer that it crashed on.

> * My bug report is sitting in the XFS bugzilla for a week without response.

I've heard negative "things" about the XFS developtment process, but
not many "things", and not from many sources, and no more than I hear
about a lot of other open source development efforts.  Those "things"
made me a little uneasy, but since I've never (to my knowledge)
experienced the issues described, and I've otherwise been very happy
with XFS, I've decided to continue to be very happy with XFS.

Also, and please don't take this the wrong way, but not having read
your bug report, I can't express any opinion on whether their ignoring
your report says good or bad things about the XFS development
process. :)

> * Grub cannot boot from XFS on AMD64.

Works fine here.  I have an Opteron and two Athlon64's all booting
grub from XFS.  (Not that it should really matter to you.  What
serious sysadmin doesn't have /boot as a separate filesystem? ;)

Now it might be the case that grub cannot boot from a *crashed* XFS
filesystem, either on AMD64 or in general, but my /boot is mounted ro
(with some nifty scripting to automate remounting during the kernel
install process), so I doubt I've ever tried to boot grub off of a
crashed XFS filesystem.  I should point out, however, that I've never
seen anything indicating that grub can replay the journal of *any* of
the journaling filesystems for which it has stage1_5 support.  On the
contrary, when they first added support for ext3, it was documented to
be incapable of replaying the journal.  Sticking with ext2 for /boot
was strongly recommended.

> * If my system crashes (while I am trying to find optimal performance
> settings), the log is often found to be corruct.
> * Then I have to boot from a different media and clear the log
> manually (when it should mount ro, imho).

Do you run iptables on this machine?  I had horrible stability issues
with a machine with the following combination:  AMD64, iptables, XFS,
LVM, RAID.  The problems manifested themselves in the XFS code.  I did
a lot of ugly reconstruction cycles.  I eliminated iptables from the
equation, and the machine has been stable ever since.  I never tried
to figure out if it was actually iptables' fault, or if it was XFS's
fault, or if it was a combination of the two (aren't monolithic
kernels great!).  Since the problem was on a production machine, I
doubt I will try to isolate it anytime soon.

> * XFS performance for file creation and deletion is horrible.

Really?  Compared to what?  That's not what the MythTV folks say.
They care about deletion in particular (big files, sometimes have to
expire an old recording before you can start a new one).  The last
data I read from them had JFS as the fastest for deletion and XFS a
close second.  That's not surprising, since SGI's main business has
always been with people who deal with video.

> So is it time to switch? I have heard that ext3 made some progress on
> the performance front, I would probably still go for reiserfs 3.6.
> 
> Or what would be your recommended file system?

I recommend XFS. :)

Brian
-- 
We Americans live in a nation where the medical-care system is second
to none in the world, unless you count maybe 25 or 30 little scuzzball
countries like Scotland that we could vaporize in seconds if we felt
like it.   -- Dave Barry, "Stay Fit & Healthy Until You're Dead"