[Gluster-users] Virtual machines and self-healing on GlusterFS v3.3
Dario Berzano
dario.berzano at cern.ch
Mon Sep 17 10:18:33 UTC 2012
Hello Pranith,
ok, I understand that each time a write operation is performed the update flag is set, and reset afterwards, when update is complete.
I really don't know GlusterFS internals, but... what if a live migration or brick failure happens *while* these updates are ongoing?
The problem is, my VMs are definitely *not* doing fine :(
My former GlusterFS configuration had only one brick, and everything went perfect. Problems started to arise as soon as we migrated to a replicated infrastructure.
I wonder if the problem is:
- our network;
- some obscure configuration internal to the VMs;
- GlusterFS;
- a combination of all the above.
Since the only thing that changed is our GlusterFS configuration, I'm "pointing the finger" to it: we've put in place such replicated configuration to avoid having single points of failures, *but* we are experiencing more failures since then! To me it seems very related to this issue, if not exactly the same problem, although on a much smaller scale:
http://gluster.org/pipermail/gluster-users/2012-September/011444.html
Of course the problem might be network-related, I am currently running tests to sort it out.
Cheers
--
: Dario Berzano
: CERN PH-SFT & Università di Torino (Italy)
: Wiki: http://newton.ph.unito.it/~berzano
: GPG: http://newton.ph.unito.it/~berzano/gpg
: Mobiles: +41 766124782 (CH), +39 3487222520 (IT)
Il giorno 17/set/2012, alle ore 10:41, Pranith Kumar Karampuri <pkarampu at redhat.com>
ha scritto:
> Dario,
> Nothing to worry then :-). It was a transient state. Every time an update is done they are marked and after the update is over they are reset. Similarly the output of 'gluster volume heal <volname> info' Keeps giving entries when these flags are set and not show any output when these flags are reset. I thought it was a persistent one. Seems like your Vm files are doing fine.
>
> Pranith.
> ----- Original Message -----
> From: "Dario Berzano" <dario.berzano at cern.ch>
> To: "Pranith Kumar Karampuri" <pkarampu at redhat.com>
> Cc: "gluster-users" <gluster-users at gluster.org>
> Sent: Monday, September 17, 2012 1:36:56 PM
> Subject: Re: [Gluster-users] Virtual machines and self-healing on GlusterFS v3.3
>
> Hi Pranith,
>
> those bricks stay on different servers connected on the same switch: the only possibility I see is that the switch went down for some reason, it is our only single point of failure. The servers themselves never went down at the same time.
>
> I do not understand however why if I run getfattr continuously:
>
> watch -n1 'getfattr -d -m . -e hex 1814/images/disk.0'
>
> I get alternating:
>
> trusted.afr.VmDir-client-0=0x000000010000000000000000
> trusted.afr.VmDir-client-1=0x000000010000000000000000
>
> and:
>
> trusted.afr.VmDir-client-0=0x000000000000000000000000
> trusted.afr.VmDir-client-1=0x000000000000000000000000
>
> This again happens with every "big" file.
>
> Does this suggest a network problem, maybe? One of the servers has 1 GbE while the other one has a faster 10 GbE, but I do not think this is enough to continuously de-synchronize the bricks...
>
> Cheers
> --
> : Dario Berzano
> : CERN PH-SFT & Università di Torino (Italy)
> : Wiki: http://newton.ph.unito.it/~berzano
> : GPG: http://newton.ph.unito.it/~berzano/gpg
> : Mobiles: +41 766124782 (CH), +39 3487222520 (IT)
>
>
>
> Il giorno 17/set/2012, alle ore 00:11, Pranith Kumar Karampuri <pkarampu at redhat.com>
> ha scritto:
>
>> 1814/images/disk.0 has pending data change log for both subvolumes. i.e. 0x00000001. This happens when both the bricks go out at the same time, while an operation is in progress. Did that happen?
>>
>> Pranith.
>>
>> ----- Original Message -----
>> From: "Dario Berzano" <dario.berzano at cern.ch>
>> To: "Pranith Kumar Karampuri" <pkarampu at redhat.com>
>> Cc: "gluster-users" <gluster-users at gluster.org>
>> Sent: Sunday, September 16, 2012 9:20:23 PM
>> Subject: Re: [Gluster-users] Virtual machines and self-healing on GlusterFS v3.3
>>
>> Ok, here's the output for 1816/images/disk.0:
>>
>> # file: bricks/VmDir01/1816/images/disk.0
>> security.selinux=0x756e636f6e66696e65645f753a6f626a6563745f723a66696c655f743a733000
>> trusted.afr.VmDir-client-0=0x000000000000000000000000
>> trusted.afr.VmDir-client-1=0x000000000000000000000000
>> trusted.gfid=0x1cef9d386f1c4424af6d95dfbcf2989b
>>
>> # file: bricks/VmDir02/1816/images/disk.0
>> security.selinux=0x756e636f6e66696e65645f753a6f626a6563745f723a66696c655f743a733000
>> trusted.afr.VmDir-client-0=0x000000000000000000000000
>> trusted.afr.VmDir-client-1=0x000000000000000000000000
>> trusted.gfid=0x1cef9d386f1c4424af6d95dfbcf2989b
>>
>> And for 1814/images/disk.0:
>>
>> # file: bricks/VmDir01/1814/images/disk.0
>> security.selinux=0x756e636f6e66696e65645f753a6f626a6563745f723a66696c655f743a733000
>> trusted.afr.VmDir-client-0=0x000000010000000000000000
>> trusted.afr.VmDir-client-1=0x000000010000000000000000
>> trusted.gfid=0xaabc0c344ccc4cfe8e2ed588dd78323b
>>
>> # file: bricks/VmDir02/1814/images/disk.0
>> security.selinux=0x756e636f6e66696e65645f753a6f626a6563745f723a66696c655f743a733000
>> trusted.afr.VmDir-client-0=0x000000010000000000000000
>> trusted.afr.VmDir-client-1=0x000000010000000000000000
>> trusted.gfid=0xaabc0c344ccc4cfe8e2ed588dd78323b
>>
>> Note that these are just two sample files, since the problem occurs with 100% of our "big" virtual machines. Here's the whole content of the GlusterFS volume along with file sizes:
>>
>> 6.3G ./1981/images/disk.0
>> 53M ./1820/images/disk.0
>> 9.7G ./1838/images/disk.0
>> 10G ./1819/images/disk.0
>> 9.2G ./1818/images/disk.0
>> 10G ./1816/images/disk.0
>> 53M ./1962/images/disk.0
>> 10G ./1814/images/disk.0
>> 6.2G ./1988/images/disk.0
>> 10G ./1817/images/disk.0
>> 53M ./1821/images/disk.0
>>
>> We currently have 11 running VMs. The "small" ones (53 MB big) have never shown any problem so far. *All* the other VMs (6 to 10 GB big) periodically show up in the output of:
>>
>> gluster volume heal VmDir info
>>
>> when there's some intense I/O occuring, disappearing immediately shortly afterwards.
>>
>> Thanks, cheers,
>> --
>> : Dario Berzano
>> : CERN PH-SFT & Università di Torino (Italy)
>> : Wiki: http://newton.ph.unito.it/~berzano
>> : GPG: http://newton.ph.unito.it/~berzano/gpg
>> : Mobiles: +41 766124782 (CH), +39 3487222520 (IT)
>>
>>
>> Il giorno 14/set/2012, alle ore 18:21, Pranith Kumar Karampuri <pkarampu at redhat.com> ha scritto:
>>
>>> Dario,
>>> Ok that confirms that it is not a split-brain. Could you post the getfattr output I requested as well?. What is the size of the VM files?.
>>>
>>> Pranith
>>> ----- Original Message -----
>>> From: "Dario Berzano" <dario.berzano at cern.ch>
>>> To: "Pranith Kumar Karampuri" <pkarampu at redhat.com>
>>> Cc: "<gluster-users at gluster.org>" <gluster-users at gluster.org>
>>> Sent: Friday, September 14, 2012 9:42:38 PM
>>> Subject: Re: [Gluster-users] Virtual machines and self-healing on GlusterFS v3.3
>>>
>>>
>>> # gluster volume heal VmDir info healed
>>>
>>>
>>> Heal operation on volume VmDir has been successful
>>>
>>>
>>> Brick one-san-01:/bricks/VmDir01
>>> Number of entries: 259
>>> Segmentation fault (core dumped)
>>>
>>>
>>> (same story for heal-failed) which seems to be exactly this bug:
>>>
>>>
>>> https://bugzilla.redhat.com/show_bug.cgi?id=836421
>>>
>>>
>>> Should I upgrade to latest QA RPMs to see what is going on?
>>>
>>>
>>> Btw, with split-brain I have no entries:
>>>
>>>
>>>
>>> Heal operation on volume VmDir has been successful
>>>
>>>
>>> Brick one-san-01:/bricks/VmDir01
>>> Number of entries: 0
>>>
>>>
>>> Brick one-san-02:/bricks/VmDir02
>>> Number of entries: 0
>>>
>>>
>>> Thank you, cheers,
>>> --
>>> : Dario Berzano
>>> : CERN PH-SFT & Università di Torino (Italy)
>>> : Wiki: http://newton.ph.unito.it/~berzano
>>> : GPG: http://newton.ph.unito.it/~berzano/gpg
>>> : Mobiles: +41 766124782 (CH), +39 3487222520 (IT)
>>>
>>>
>>>
>>>
>>> Il giorno 14/set/2012, alle ore 17:16, Pranith Kumar Karampuri < pkarampu at redhat.com >
>>> ha scritto:
>>>
>>>
>>> hi Dario,
>>> Could you post the output of the following commands:
>>> gluster volume heal VmDir info healed
>>> gluster volume heal VmDir info split-brain
>>>
>>> Also provide the output of 'getfattr -d -m . -e hex' On both the bricks for the two files listed in the output of 'gluster volume heal VmDir info'
>>>
>>> Pranith.
>>>
>>> ----- Original Message -----
>>> From: "Dario Berzano" < dario.berzano at cern.ch >
>>> To: gluster-users at gluster.org
>>> Sent: Friday, September 14, 2012 6:57:32 PM
>>> Subject: [Gluster-users] Virtual machines and self-healing on GlusterFS v3.3
>>>
>>>
>>>
>>> Hello,
>>>
>>>
>>> in our computing centre we have an infrastructure with a GlusterFS volume made of two bricks in replicated mode:
>>>
>>>
>>>
>>>
>>>
>>> Volume Name: VmDir
>>> Type: Replicate
>>> Volume ID: 9aab85df-505c-460a-9e5b-381b1bf3c030
>>> Status: Started
>>> Number of Bricks: 1 x 2 = 2
>>> Transport-type: tcp
>>> Bricks:
>>> Brick1: one-san-01:/bricks/VmDir01
>>> Brick2: one-san-02:/bricks/VmDir02
>>>
>>>
>>>
>>>
>>> We are using this volume to store running images of some KVM virtual machines and thought we could benefit from the replicated storage in order to achieve more robustness as well as the ability to live-migrate VMs.
>>>
>>>
>>> Our GlusterFS volume VmDir is mounted on several (three at the moment) hypervisors.
>>>
>>>
>>> However, in many cases (but it is difficult to reproduce: best way is to stress VM I/O), either when one brick becomes unavailable for some reason, or when we perform live migrations, virtual machines decide to remount filesystems from their virtual disks in read-only. At the same time, on the hypervisors mounting the GlusterFS partitions, we spot some kernel messages like:
>>>
>>>
>>>
>>>
>>> INFO: task kvm:13560 blocked for more than 120 seconds.
>>>
>>>
>>>
>>>
>>> By googling it I have found some "workarounds" to mitigate this problem, like mounting disks within virtual machines with barrier=0:
>>>
>>>
>>> http://invalidlogic.com/2012/04/28/ubuntu-precise-on-xenserver-disk-errors/
>>>
>>>
>>> but I actually fear to damage my virtual machine disks by doing such a thing!
>>>
>>>
>>> AFAIK from GlusterFS v3.3 self-healing should be performed server-side (and no self-healing at all is performed on the clients and by granularly locking big files). When I connect to my GlusterFS pool, if I monitor the self-healing status continuously:
>>>
>>>
>>> watch -n1 'gluster volume heal VmDir info'
>>>
>>>
>>> I obtain an output like:
>>>
>>>
>>>
>>>
>>>
>>> Heal operation on volume VmDir has been successful
>>>
>>>
>>> Brick one-san-01:/bricks/VmDir01
>>> Number of entries: 2
>>> /1814/images/disk.0
>>> /1816/images/disk.0
>>>
>>>
>>> Brick one-san-02:/bricks/VmDir02
>>> Number of entries: 2
>>> /1816/images/disk.0
>>> /1814/images/disk.0
>>>
>>>
>>>
>>>
>>> with a list of virtual machine disks healed by GlusterFS. Those and other files continuously appear and disappear from the list.
>>>
>>>
>>> This is a behavior I don't understand at all: does this mean that those files continuously get corrupted and healed, and self-healing is just a natural part of the replication process?! Or some kind of corruption is actually happening on our virtual disks for some reason? Is this related to the "remount readonly" problem?
>>>
>>>
>>> A more general question maybe would be: is GlusterFS v3.3 ready for storing running virtual machines (and is there some special configuration option needed on the volumes and clients for that)?
>>>
>>> Thank you in advance for shedding some light...
>>>
>>>
>>> Regards,
>>>
>>> --
>>> : Dario Berzano
>>> : CERN PH-SFT & Università di Torino (Italy)
>>> : Wiki: http://newton.ph.unito.it/~berzano
>>> : GPG: http://newton.ph.unito.it/~berzano/gpg
>>> _______________________________________________
>>> Gluster-users mailing list
>>> Gluster-users at gluster.org
>>> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20120917/c33f7606/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4050 bytes
Desc: not available
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20120917/c33f7606/attachment.p7s>
More information about the Gluster-users
mailing list