[CentOS] XFS-filesystem corrupted by defragmentation Was: Performance problems with XFS on Centos 5.4

Tue Apr 13 18:06:45 UTC 2010
Bernhard Gschaider <bgschaid_lists at ice-sf.at>

Just to close this thread and remove any doubt that it might have
raised about XFS: the problem was a PEBCAB[1] 
It was pointed out to me on the XFS-list that the device I used for
xfs_db was inconsistent with the info from xfs_info (I was blindly
copying the device from the output of df)

Footnotes: 
[1]  PEBCAB: Problem exists between chair and keyboard


>>>>> On Tue, 13 Apr 2010 11:54:53 +0200
>>>>> "BG" == Bernhard Gschaider <bgschaid_lists at ice-sf.at> wrote:

    BG> Before I'd try to defragment my whole filesystem (see attached
    BG> mail for whole story) I figured "Let's try it on some file".

    BG> So I did

    >> xfs_bmap /raid/Temp/someDiskimage.iso
    BG> [output shows 101 extents and 1 hole]

    BG> Then I defragmented the file
    >> xfs_fsr /raid/Temp/someDiskimage.iso
    BG> extents before:101 after:3 DONE

    >> xfs_bmap /raid/Temp/someDiskimage.iso
    BG> [output shows 3 extents and 1 hole]

    BG> and now comes the bummer: i wanted to check the fragmentation
    BG> of the whole filesystem (just for checking):

    >> xfs_db -r /dev/mapper/VolGroup00-LogVol04
    BG> xfs_db: unexpected XFS SB magic number 0x00000000 xfs_db: read
    BG> failed: Invalid argument xfs_db: data size check failed
    BG> cache_node_purge: refcount was 1, not zero (node=0x2a25c20)
    BG> xfs_db: cannot read root inode (22)

    BG> THAT output was definitly not there when I did this the last
    BG> time and therefor the new fragmentation does not make me happy
    BG> either

    xfs_db> frag
    BG> actual 0, ideal 0, fragmentation factor 0.00%

    BG> The file-system is still mounted and working and I don't dare
    BG> to do anything about it (am in a mild state of panic) because
    BG> I think it might not come back if I do.

    BG> Any suggestions most welcome (am googling myself before I do
    BG> anything about it).

    BG> I swear to god: I did not do anything else with the
    BG> xfs_*-commands than the stuff mentioned above

    BG> Bernhard

    BG> From: Bernhard Gschaider <bgschaid_lists at ice-sf.at> Subject:
    BG> Re: [CentOS] Performance problems with XFS on Centos 5.4 To:
    BG> CentOS mailing list <centos at centos.org> Date: Mon, 12 Apr 2010
    BG> 18:22:24 +0200 Organization: ICE Stroemungsforschung Reply-To:
    BG> CentOS mailing list <centos at centos.org>


>>>>> On Fri, 9 Apr 2010 10:59:02 -0400
>>>>> "RW" == Ross Walker <rswwalker at gmail.com> wrote:

    RW> On Apr 9, 2010, at 9:59 AM, Bernhard Gschaider
    RW> <bgschaid_lists at ice-
    sf.at> wrote:

    >>> Hi!
    >>> 
    >>> During the last weeks I experienced some performance problems
    >>> with a large file-system on XFS basis. Sometimes for instance
    >>> ls is painfully. Immidiatly afterwards ls on the same
    >>> directory is immidiate. I used strace on this ls and found
    >>> that during the first ls the lstat-calls need approx 0.02s
    >>> each while during the second ls the are two orders of
    >>> magnitude faster.
    >>> 
    >>> Googling around I stumbled upon some messages similar like
    >>> this
    >>> 
    >>> http://www.opensubscriber.com/message/linux-xfs@oss.sgi.com/1355060.html
    >>> 
    >>> which have in common a) they're from around 2006 b) they
    >>> suggest to increase a mount-option ihashsize. This mount
    >>> option is listed as deprecated in the current kernel-doc
    >>> 
    >>> So my question: does anyone have experience with that kind of
    >>> performance problem? Do you think it is a XFS problem or are
    >>> there some other tuning parameters in the kernel that could be
    >>> modified for instance via /proc?
    >>> 
    >>> The reason why I'm asking here is that it is a production
    >>> file-system so I would be very unpopular if I experiment too
    >>> much (a couple of reboots is OK ;) )
    >>> 
    >>> Bernhard
    >>> 
    >>> PS: the situation got worse during the last weeks when the
    >>> file-system increased in size, so the option that some kind of
    >>> buffer now is too small and I'm experiencing some kind of
    >>> thrashing seems very likely to me

    RW> Are you defragging the file system regularly?

    BG> Uups. Never occured to me ("Fragmentation is soooo Windoze")
    BG> Had a look:

    xfs_db> frag
    BG> actual 6349355, ideal 4865683, fragmentation factor 23.37%

    BG> This seems significant.

    RW> How much memory do you have in the system and how big is the
    RW> file system?

    BG> Memory on the system is 4Gig (2 DualCore Xenons). The
    BG> filesystem is 3.5 TB of which 740 Gig are used. Which is the
    BG> maximum amount used during the one year that the filesystem is
    BG> being used (that is why the high fragmentation amazes me)

    RW> What are the XFS parameters for the file system?

    BG> Is this sufficent?

    BG> % xfs_info /raid meta-data=/dev/VolGroup00/LogVol05 isize=256
    BG> agcount=32, agsize=29434880 blks = sectsz=512 attr=0 data =
    BG> bsize=4096 blocks=941916160, imaxpct=25 = sunit=0 swidth=0
    BG> blks, unwritten=1 naming =version 2 bsize=4096 log =internal
    BG> bsize=4096 blocks=32768, version=1 = sectsz=512 sunit=0 blks,
    BG> lazy-count=0 realtime =none extsz=4096 blocks=0, rtextents=0


    RW> What is the storage setup?

    BG> The filesystem is on a LVM-Volume which sits on a RAID 5
    BG> (Hardware RAID) drive

    RW> Need the info.

    BG> So the way to go forward would be using xfs_fsr on that
    BG> drive. I read some horror stories about lost files, are these
    BG> to be taken seriously (I mean they were in some Ubuntu forums
    BG> ;) )

    BG> Any other thoughts on parameters?

    BG> Thanks for your time

    BG> Bernhard _______________________________________________
    BG> CentOS mailing list CentOS at centos.org
    BG> http://lists.centos.org/mailman/listinfo/centos

    BG> ----------

    BG> _______________________________________________ CentOS mailing
    BG> list CentOS at centos.org
    BG> http://lists.centos.org/mailman/listinfo/centos


-- 
---------------------------------------------------------------------------
DI Bernhard F.W. Gschaider
---------------------------------------------------------------------------
EMail: 	Bernhard.Gschaider at ice-sf.at
WWW  : www.ice-sf.at
Jabber : bgschaid at jabber.org
Tel:	+43(3842)98282-42	Fax:	+43(3842)98282-02
---------------------------------------------------------------------------