On 01/14/2013 10:13 AM, Doron Fediuck wrote:
------------------------------------------------------------------------
*From: *"Alexandru Vladulescu" <[email protected]>
*To: *"Doron Fediuck" <[email protected]>
*Cc: *"users" <[email protected]>
*Sent: *Sunday, January 13, 2013 9:49:25 PM
*Subject: *Re: [Users] Testing High Availability and Power outages
Dear Doron,
I had the case retested now and I am writing you the results.
Furthermore, if this information should be useful for you, my
network setup is the following: 2 Layer 2 (Zyxel es2108-g &
ES2200-8) switches configured with 2 VLANs ( 1 inside backbone
network -- added br0 to Ovirt ; 1 outside network -- running on
ovirtmgmt interface for Internet traffic to VMs). The backbone
switch is a gigabit capable one, and each host runs on jumbo frame
setup. There is one more firewall server that routes the subnets
through trunking port and VLAN configuration. The Ovirt software
has been setup with backbone network subnet.
As you could guess the network infrastructure is not the problem here.
The test case was the same as described before:
1. Vm running on Hyper01, none on Hyper02. Host had configured the
High Available check box.
2. Hand power off of Hyper01 from power network (no soft/manual
shutdown).
3. After a while, Ovirt marks the Hyper01 as Non Responsive
4. Manually clicked on Confirm host reboot and the VM starts after
Ovirt's manual fence to Hyper01 on Hyper02 host.
I have provided engine log attached. The Confirm Host reboot was
done at precise time of 21:31:45 On the cluster section, in Ovirt,
I did try changing the "Resilience Policy" attribute from "Migrate
Virtual Machines" to "Migrate only High Available Virtual
Machines" but with the same results.
As I am guessing from the engine log the Node Controller sees the
Hyper01 node as it has a "network fault" no route to host,
although this was shut down.
Is this supposed to be the default behavior in this case, as the
scenario might overlap with a real case of network outage.
My Regards,
Alex.
On 01/13/2013 10:54 AM, Doron Fediuck wrote:
------------------------------------------------------------------------
*From: *"Alexandru Vladulescu" <[email protected]>
*To: *"Doron Fediuck" <[email protected]>
*Cc: *"users" <[email protected]>
*Sent: *Sunday, January 13, 2013 10:46:41 AM
*Subject: *Re: [Users] Testing High Availability and Power
outages
Dear Doron,
I haven't collected the logs from the tests, but I would
gladly re-do the case and get back to you asap.
This feature is the main reason of which I have chosen to
go with Ovirt in the first place, besides other virt
environments.
Could you please inform me what logs should I be focusing
on, besides the engine log; vdsm maybe or other relevant logs?
Regards,
Alex
--
Sent from phone.
On 13.01.2013, at 09:56, Doron Fediuck
<[email protected] <mailto:[email protected]>> wrote:
------------------------------------------------------------------------
*From: *"Alexandru Vladulescu"
<[email protected]
<mailto:[email protected]>>
*To: *"users" <[email protected]
<mailto:[email protected]>>
*Sent: *Friday, January 11, 2013 2:47:38 PM
*Subject: *[Users] Testing High Availability and
Power outages
Hi,
Today, I started testing on my Ovirt 3.1
installation (from dreyou repos) running on 3 x
Centos 6.3 hypervisors the High Availability
features and the fence mechanism.
As yesterday, I have reported in a previous email
thread, that the migration priority queue cannot
be increased (bug) in this current version, I
decided to test what the official documentation
says about the High Availability cases.
This will be a disaster case scenarios to suffer
from if one hypervisor has a power outage/hardware
problem and the VMs running on it are not
migrating on other spare resources.
In the official documenation from ovirt.org
<http://ovirt.org> it is quoted the following:
/High availability /
//
/Allows critical VMs to be restarted on another
host in the event of hardware failure with three
levels of priority, taking into account resiliency
policy. /
//
* /Resiliency policy to control high
availability VMs at the cluster level. /
* /Supports application-level high availability
with supported fencing agents. /
As well as in the Architecture description:
/High Availability - restart guest VMs from failed
hosts automatically on other hosts/
So the testing went like this -- One VM running a
linux box, having the check box "High Available"
and "Priority for Run/Migration queue:" set to
Low. On Host we have the check box to "Any Host in
Cluster", without "Allow VM migration only upon
Admin specific request" checked.
My environment:
Configuration : 2 x Hypervisors (same
cluster/hardware configuration) ; 1 x Hypervisor +
acting as a NAS (NFS) server (different
cluster/hardware configuration)
Actions: Went and cut-off the power from one of
the hypervisors from the 2 node clusters, while
the VM was running on. This would translate to a
power outage.
Results: The hypervisor node that suffered from
the outage is showing in Hosts tab as Non
Responsive on Status, and the VM has a question
mark and cannot be powered off or nothing
(therefore it's stuck).
In the Log console in GUI, I get:
Host Hyper01 is non-responsive.
VM Web-Frontend01 was set to the Unknown status.
There is nothing I could I could do besides
clicking on the Hyper01 "Confirm Host as been
rebooted", afterwards the VM starts on the Hyper02
with a cold reboot of the VM.
The Log console changes to:
Vm Web-Frontend01 was shut down due to Hyper01
host reboot or manual fence
All VMs' status on Non-Responsive Host Hyper01
were changed to 'Down' by admin@internal
Manual fencing for host Hyper01 was started.
VM Web-Frontend01 was restarted on Host Hyper02
I would like you approach on this problem, reading
the documentation & features pages on the official
website, I suppose that this would have been an
automatically mechanism working on some sort of a
vdsm & engine fencing action. Am I missing
something regarding it ?
Thank you for your patience reading this.
Regards,
Alex.
_______________________________________________
Users mailing list
[email protected] <mailto:[email protected]>
http://lists.ovirt.org/mailman/listinfo/users
Hi Alex,
Can you share with us the engine's log from the
relevant time period?
Doron
Hi Alex,
engine log is the important one, as it will indicate on the
decision making process.
VDSM logs should be kept in case something is unclear, but I
suggest we begin with
engine.log.
Hi Alex,in tab, rig
In order to have HA working in host level (which is what you're
testing now) you need to
configure power management to each of the relevant hosts (Go to Hosts
maht click a host
and choose edit. Now select the Power management tab and you'll see
it). In the details you
gave us it's not clear how you defined Power management for your
hosts, so I can only assume
it's not defined properly.
The reason for this necessity is that we cannot resume a VM on a
different host before we
verified the original hosts status. If, for example the VM is still
running on the original
host and we lost network connectivity to it, we're in a risk of
running the same VM on 2 different
hosts at the same time which will corrupt its disk(s). So the only way
to prevent it, is
rebooting the original host which will ensure the VM is not running
there. We call the reboot
procedure fencing, and if you'll check your logs you'll be able to see:
2013-01-13 21:29:42,380 ERROR
[org.ovirt.engine.core.bll.VdsNotRespondingTreatmentCommand]
(pool-3-thread-44) [a1803d1] Failed to run Fence script on
vds:Hyper01, VMs moved to UnKnown instead.
So the only way for you to handle it, is to confirm host was rebooted
(as you did), which will
allow resuming the VM on a different host.
Doron
Hi Doron,
Regarding your reply I don't have such fence mechanism through IMM or
iLO interface as the hardware that I am using doesn't support such IPMI
technology. Seeing your response makes me consider the option of really
getting an add-on card that will be able to do the basic reboot,
restart, reset functions for our hardware.
Thank you very much for your advice on this.
Alex
_______________________________________________
Users mailing list
[email protected]
http://lists.ovirt.org/mailman/listinfo/users
