On Mon, Apr 23, 2018 at 8:06 PM, Michal Skrivanek <michal.skrivanek@redhat.com> wrote:


On 23 Apr 2018, at 10:52, Daniel Menzel <daniel.menzel@hhi.fraunhofer.de> wrote:

Hi Michal,

in your last mail you wrote, that the values can be turned down - how can this be done?

​H
​AFAIK , there is no point in changing fencing vdc_options values in that case (assuming no kdump is configured here ...)

The Fencing mechanism ​
 
​is putting the host in "connecting" state for a grace period that depends on its number of running VMs and if it serves as APM or not
While the host became non-responding , we first try to do a soft-fence (restart VDSM via ssh) , this will also take time
After that point , if soft-fence is failing , the host will be reboot via the fencing script and the time it takes is totally depending ​on the host
If you have something to look at , it is your host reboot time and try to improve it, if the host will reboot faster, the whole process will take less time ...

Regards

Eli



Best
Daniel


On 12.04.2018 20:29, Michal Skrivanek wrote:


On 12 Apr 2018, at 13:13, Daniel Menzel <daniel.menzel@hhi.fraunhofer.de> wrote:

Hi there,

does anyone have an idea how to decrease a virtual machine's downtime?

Best
Daniel


On 06.04.2018 13:34, Daniel Menzel wrote:

Hi Michal,



Hi Daniel,
adding Martin to review fencing behavior

(sorry for misspelling your name in my first mail).



that’s not the reason I’m replying late!:-))

The settings for the VMs are the following (oVirt 4.2):

  1. HA checkbox enabled of course
  2. "Target Storage Domain for VM Lease" -> left empty

if you need faster reactions then try to use VM Leases as well, it won’t make a difference in this case but will help in case of network issues. E.g. if you use iSCSI and the storage connection breaks while host connection still works it would restart the VM in about 80s; otherwise it would take >5 mins. 
  1. "Resume Behavior" -> AUTO_RESUME
  2. Priority for Migration -> High
  3. "Watchdog Model" -> No-Watchdog

For testing we did not kill any VM but the host. So basically we simulated an instantaneous crash by manually turning the machine off via IPMI-Interface (not via operating system!) and ping the guest(s). What happens then?

  1. 2-3 seconds after the we press the host's shutdown button we lose ping contact to the VM(s).
  2. After another 20s oVirt changes the host's status to "connecting", the VM's status is set to a question mark.
  3. After ~1:30 the host is flagged to "non responsive”

that sounds about right. Now fencing action should have been initiated, if you can share the engine logs we can confirm that. IIRC we first try soft fencing - try to ssh to that host, that might take some time to time out I guess. Martin?

  1. After ~2:10 the host's reboot is initiated by oVirt, 5-10s later the guest is back online.

So, there seems to be one mistake I made in the first mail: The downtime is "only" 2.5min. But still I think this time can be decreased as for some services it is still quite a long time.



these values can be tuned down, but then you may be more susceptible to fencing power cycling a host in case of shorter network outages. It may be ok…depending on your requirements.

Best
Daniel


On 06.04.2018 12:49, Michal Skrivanek wrote:
On 6 Apr 2018, at 12:45, Daniel Menzel <daniel.menzel@hhi.fraunhofer.de> wrote:

Hi Michael,
thanks for your mail. Sorry, I forgot to write that. Yes, we have power management and fencing enabled on all hosts. We also tested this and found out that it works perfectly. So this cannot be the reason I guess.
Hi Daniel,
ok, then it’s worth looking into details. Can you describe in more detail what happens? What exact settings you’re using for such VM? Are you killing the HE VM or other VMs or both? Would be good to narrow it down a bit and then review the exact flow

Thanks,
michal

Daniel



On 06.04.2018 11:11, Michal Skrivanek wrote:
On 4 Apr 2018, at 15:36, Daniel Menzel <daniel.menzel@hhi.fraunhofer.de> wrote:

Hello,

we're successfully using a setup with 4 Nodes and a replicated Gluster for storage. The engine is self hosted. What we're dealing with at the moment is the high availability: If a node fails (for example simulated by a forced power loss) the engine comes back up online withing ~2min. But guests (having the HA option enabled) come back online only after a very long grace time of ~5min. As we have a reliable network (40 GbE) and reliable servers I think that the default grace times are way too high for us - is there any possibility to change those values?
And do you have Power Management(iLO, iDRAC,etc) configured for your hosts? Otherwise we have to resort to relatively long timeouts to make sure the host is really dead
Thanks,
michal
Thanks in advance!
Daniel

_______________________________________________
Users mailing list
Users@ovirt.org
http://lists.ovirt.org/mailman/listinfo/users





_______________________________________________
Users mailing list
Users@ovirt.org
http://lists.ovirt.org/mailman/listinfo/users

_______________________________________________
Users mailing list
Users@ovirt.org
http://lists.ovirt.org/mailman/listinfo/users