thus Ross Walker spake:
On Dec 9, 2009, at 8:05 AM, Timo Schoeler timo.schoeler@riscworks.net wrote:
Hi list,
during the last days there was a discussion going on about the stability of XFS; though I myself used XFS heavily and didn't run into issues yet, I'd like to ask something *before* we create our next generation data storage backend...
Les Mikesell wrote in [0] about issues in the combination of XFS and LVM -- however, it was being discussed in context of using 32bit kernels.
What I specifically need is to run XFS (or something similar, I am *not* forced to use XFS, but it was my preference for some years now, and I didn't have any issues with it yet) on top of LVM to be able to create snapshots. We're talking about several file systems of a size at about 4TiByte each.
On another place [1] I read that there were issues with that.
Can anyone shed some light on this? Would be very appreciated.
There is no problem if it is done on x86_64 with it's 8k stack frames, but on i386 with it's 4k stack frames you could run into a stack overflow when doing it on top of stackable block devices (md raid, lvm, drbd, etc).
Also since the current LVM on CentOS doesn't support barriers (next release I believe) journalling isn't safe on LVM unless you are using a storage controller with BBU write-back cache.
I have heard anyways that the current implementation of barriers isn't very performant and doesn't take into consideration controllers with BBU cache, so most people will end up mounting with nobarriers which just means they are in the same boat as they are now. Better make sure your machine is bullet proof as a power outage or a kernel panic can spell disaster for XFS (or any other file system really).
It is better to invest in a good hardware RAID controller until the whole barriers stuff is ironed out. It should really perform better then it does.
Thanks for your detailed explanation, that really clears things up; however, I was intending to build a software RAID10 as we had really not so good experiences on hw RAID controllers int the past (for all kinds of phenomena).
Would barriering here still be a problem then?
Timo
-Ross