[ovirt-users] Glusterfs HA with Ovirt

Brad House brad at monetra.com
Fri Jul 4 13:27:37 UTC 2014


I was never able to achieve a stable system that could survive the loss of a
single node with glusterfs.  I attempted to use replica 2 across 3 nodes (which
required 2 bricks per node as the number of bricks must be a multiple of the
replica, and you have to order them so the brick pairs span servers).  I enabled
server-side quorum, but found out later that client side quorum is based on
'sub volumes', which means that with a single node failure on replica 2, even
though there were 3 nodes, it would go into a readonly state.

After disabling client-side quorum (but keeping server side quorum) I thought the
issue was fixed, but every once in a while, rebooting one of the nodes (after ensuring
gluster was healed) would lead to i/o errors on the VM guest and essentially make
it so it needed to be rebooted (which was successful and everything worked after
even before bringing the downed node back up).  My nodes were all combined glusterfs
and ovirt nodes. I tried using both 'localhost' on the nodes as well as using a
keepalived VIP.

Its possible my issues were all due to client-side quorum not being enabled,
but that would require replica 3 to be able to survive a single node failure, but
I never pursued testing that theory.  Also, heal times seemed a bit long for
a single idle VM, it would consume 2 full cores of CPU for about 5 minutes for
healing a single idle VM (granted, I was testing on a 1Gbps network, but that
doesn't explain the CPU usage).

-Brad

On 7/4/14 1:29 AM, Andrew Lau wrote:
> As long as all your compute nodes are part of the gluster peer,
> localhost will work.
> Just remember, gluster will connect to any server, so even if you
> mount as localhost:/ it could be accessing the storage from another
> host in the gluster peer group.
>
>
> On Fri, Jul 4, 2014 at 3:26 PM, Punit Dambiwal <hypunit at gmail.com> wrote:
>> Hi Andrew,
>>
>> Yes..both on the same node...but i have 4 nodes of this type in the same
>> cluster....So it should work or not ??
>>
>> 1. 4 physical nodes with 12 bricks each(distributed replicated)...
>> 2. The same all 4 nodes use for the compute purpose also...
>>
>> Do i still require the VIP...or not ?? because i tested even the mount point
>> node goes down...the VM will not pause and not affect...
>>
>>
>> On Fri, Jul 4, 2014 at 1:18 PM, Andrew Lau <andrew at andrewklau.com> wrote:
>>>
>>> Or just localhost as your computer and storage are on the same box.
>>>
>>>
>>> On Fri, Jul 4, 2014 at 2:48 PM, Punit Dambiwal <hypunit at gmail.com> wrote:
>>>> Hi Andrew,
>>>>
>>>> Thanks for the update....that means HA can not work without VIP in the
>>>> gluster,so better to use the glusterfs with the VIP to take over the
>>>> ip...in
>>>> case of any storage node failure...
>>>>
>>>>
>>>> On Fri, Jul 4, 2014 at 12:35 PM, Andrew Lau <andrew at andrewklau.com>
>>>> wrote:
>>>>>
>>>>> Don't forget to take into consideration quroum, that's something
>>>>> people often forget
>>>>>
>>>>> The reason you're having the current happen, is gluster only uses the
>>>>> initial IP address to get the volume details. After that it'll connect
>>>>> directly to ONE of the servers, so with your 2 storage server case,
>>>>> 50% chance it won't go to paused state.
>>>>>
>>>>> For the VIP, you could consider CTDB or keepelived, or even just using
>>>>> localhost (as your storage and compute are all on the same machine).
>>>>> For CTDB, checkout
>>>>> http://community.redhat.com/blog/2014/05/ovirt-3-4-glusterized/
>>>>>
>>>>> I have a BZ open regarding gluster VMs going into paused state and not
>>>>> being resumable, so it's something you should also consider. My case,
>>>>> switch dies, gluster volume goes away, VMs go into paused state but
>>>>> can't be resumed. If you lose one server out of a cluster is a
>>>>> different story though.
>>>>> https://bugzilla.redhat.com/show_bug.cgi?id=1058300
>>>>>
>>>>> HTH
>>>>>
>>>>> On Fri, Jul 4, 2014 at 11:48 AM, Punit Dambiwal <hypunit at gmail.com>
>>>>> wrote:
>>>>>> Hi,
>>>>>>
>>>>>> Thanks...can you suggest me any good how to/article for the glusterfs
>>>>>> with
>>>>>> ovirt...
>>>>>>
>>>>>> One strange thing is if i will try both (compute & storage) on the
>>>>>> same
>>>>>> node...the below quote not happen....
>>>>>>
>>>>>> ---------------------
>>>>>>
>>>>>> Right now, if 10.10.10.2 goes away, all your gluster mounts go away
>>>>>> and
>>>>>> your
>>>>>> VMs get paused because the hypervisors can’t access the storage. Your
>>>>>> gluster storage is still fine, but ovirt can’t talk to it because
>>>>>> 10.10.10.2
>>>>>> isn’t there.
>>>>>> ---------------------
>>>>>>
>>>>>> Even the 10.10.10.2 goes down...i can still access the gluster mounts
>>>>>> and no
>>>>>> VM pause....i can access the VM via ssh...no connection
>>>>>> failure.....the
>>>>>> connection drop only in case of SPM goes down and the another node
>>>>>> will
>>>>>> elect as SPM(All the running VM's pause in this condition).
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Fri, Jul 4, 2014 at 4:12 AM, Darrell Budic
>>>>>> <darrell.budic at zenfire.com>
>>>>>> wrote:
>>>>>>>
>>>>>>> You need to setup a virtual IP to use as the mount point, most
>>>>>>> people
>>>>>>> use
>>>>>>> keepalived to provide a virtual ip via vrrp for this. Setup
>>>>>>> something
>>>>>>> like
>>>>>>> 10.10.10.10 and use that for your mounts.
>>>>>>>
>>>>>>> Right now, if 10.10.10.2 goes away, all your gluster mounts go away
>>>>>>> and
>>>>>>> your VMs get paused because the hypervisors can’t access the
>>>>>>> storage.
>>>>>>> Your
>>>>>>> gluster storage is still fine, but ovirt can’t talk to it because
>>>>>>> 10.10.10.2
>>>>>>> isn’t there.
>>>>>>>
>>>>>>> If the SPM goes down, it the other hypervisor hosts will elect a new
>>>>>>> one
>>>>>>> (under control of the ovirt engine).
>>>>>>>
>>>>>>> Same scenarios if storage & compute are on the same server, you
>>>>>>> still
>>>>>>> need
>>>>>>> a vip address for the storage portion to serve as the mount point so
>>>>>>> it’s
>>>>>>> not dependent on any one server.
>>>>>>>
>>>>>>> -Darrell
>>>>>>>
>>>>>>> On Jul 3, 2014, at 1:14 AM, Punit Dambiwal <hypunit at gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> I have some HA related concern about glusterfs with Ovirt...let say
>>>>>>> i
>>>>>>> have
>>>>>>> 4 storage node with gluster bricks as below :-
>>>>>>>
>>>>>>> 1. 10.10.10.1 to 10.10.10.4 with 2 bricks each and i have
>>>>>>> distributed
>>>>>>> replicated architecture...
>>>>>>> 2. Now attached this gluster storge to ovrit-engine with the
>>>>>>> following
>>>>>>> mount point 10.10.10.2/vol1
>>>>>>> 3. In my cluster i have 3 hypervisior hosts (10.10.10.5 to
>>>>>>> 10.10.10.7)
>>>>>>> SPM
>>>>>>> is on 10.10.10.5...
>>>>>>> 4. What happen if 10.10.10.2 will goes down.....can hypervisior host
>>>>>>> can
>>>>>>> still access the storage ??
>>>>>>> 5. What happen if SPM goes down ???
>>>>>>>
>>>>>>> Note :- What happen for point 4 &5 ,If storage and Compute both
>>>>>>> working
>>>>>>> on
>>>>>>> the same server.
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Punit
>>>>>>> _______________________________________________
>>>>>>> Users mailing list
>>>>>>> Users at ovirt.org
>>>>>>> http://lists.ovirt.org/mailman/listinfo/users
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Users mailing list
>>>>>> Users at ovirt.org
>>>>>> http://lists.ovirt.org/mailman/listinfo/users
>>>>>>
>>>>
>>>>
>>
>>
> _______________________________________________
> Users mailing list
> Users at ovirt.org
> http://lists.ovirt.org/mailman/listinfo/users
>



More information about the Users mailing list