On 28 Apr 2016, at 14:39, Joost(a)familiealbers.nl wrote:
Ok that sounds promising. Does it require 3.6 at minimum?
yes, the 3.6 requires far less network bandwidth during stable conditions (nothing’s going
on with VM status)
We already handle starting vms using vdsm in case the network from
hypervisors to engine is down.
that’s…well, a bit tricky to do. Can you share how exactly? What compromises did you do?
As well as disabling fencingmto avoid a sloppy network to try and
make changes. I would like to reduce the comms between hosts and engine. Are there any
ideas you would have about that.
fencing sometimes does crazy things indeed. There were also quite a few enhancements in
3.5/3.6, but iirc they are not default so you’d need to enable them (e.g. skipping fencing
when storage lease is active)
there really shouldn’t be much else
some parameters might help in unstable conditions, I guess increasing
vdsHeartbeatInSeconds would make a big difference in general at expense of slower
detection of network outages(mostly relevant for fencing which you don’t use so it
shouldn’t matter to you)
Verstuurd vanaf mijn iPhone
> Op 28 apr. 2016 om 13:55 heeft Michal Skrivanek <michal.skrivanek(a)redhat.com>
het volgende geschreven:
>
>
>>> On 26 Apr 2016, at 16:48, Martin Sivak <msivak(a)redhat.com> wrote:
>>>
>>> @awels: to add another layer of indirection via a dedicated
>>> hosted-engine per outlet seems a little much. we are talking about 500 *
>>> 4GB RAM at least in this example, so 2 TB RAM just for management
>>> purposes, if you follow engine hardware recommendations?
>>
>> I would not go that far. Creating zones per continent (for example)
>> might be enough.
>>
>>> At least RHEV states in the documentation you support up to 200 hosts
>>> per cluster alone.
>>
>> The default configuration seems to only allow 250 hosts per datacenter.
>>
>> # engine-config -g MaxNumberOfHostsInStoragePool
>> MaxNumberOfHostsInStoragePool: 250 version: general
>
> yep, but that liit is there because within a DC there is a lot of assumption for
flawless fast enough communication, the most problematic is that all hosts need to access
the same storage and the monitoring gets expensive then.
> This is a different situation with separate DCs, there’s no cross-DC communication.
> I would guess many DCs work great actually.
>
> Too many hosts and VMs in total might be an issue, but since the last official
updates there were a lot of changes. E.g. in stable state due to VM status events
introduced in 3.6 the traffic required between each host and engine is much lower.
> I would not be so afraid of thousands anymore, but of course YMMV
>
>>
>> --
>> Martin Sivak
>> SLA / oVirt
>>
>>> On Tue, Apr 26, 2016 at 4:03 PM, Sven Kieske <svenkieske(a)gmail.com>
wrote:
>>>> On 26.04.2016 14:46, Martin Sivak wrote:
>>>> I think that 1000 hosts per engine is a bit over what we recommend
>>>> (and support). The fact that all of them are going to be remote might
>>>> not be ideal either. The engine assumes the network connection to all
>>>> hosts is almost flawless and the necessary routing and distance to
>>>> your hosts might not play nice with (for example) the fencing logic.
>>>
>>> Hi,
>>>
>>> this seems a little surprising.
>>>
>>> At least RHEV states in the documentation you support up to 200 hosts
>>> per cluster alone.
>>>
>>> There are no documented maxima for clusters or datacenters though.
>>>
>>> @awels: to add another layer of indirection via a dedicated
>>> hosted-engine per outlet seems a little much. we are talking about 500 *
>>> 4GB RAM at least in this example, so 2 TB RAM just for management
>>> purposes, if you follow engine hardware recommendations?
>
> yeah. currently the added layer of manageiq with HEs everywhere is not that helpful
for this particular case. Still, a per-continent split or per-low-latency-area might not
be a bad idea.
> I can imagine with a bit more tolerant timeouts and refreshes it might work well,
with incidents/disconnects being isolated within a DC
>
>>>
>>> But I agree, ovirt does not handle unstable or remote connections that
>
> right. but most of that is again per-DC. You can’t do much cross-DC though (e.g.
sharing a template is a pain)
>
> Thanks
> michal
>
>>> well, so you might be better of with hundredths of remote engines, but
>>> it seems to be a nightmare to manage, even if you automate everything.
>>>
>>> My personal experience is, that ovirt does scale at least until about
>>> 30-50 DCs managed by a single engine, but that setup was also on a LAN
>>> (but I would say it could scale well beyond these numbers, at least on a
>>> LAN).
>>>
>>> HTH
>>>
>>> Sven
>>>
>>>
>>> _______________________________________________
>>> Users mailing list
>>> Users(a)ovirt.org
>>>
http://lists.ovirt.org/mailman/listinfo/users
>> _______________________________________________
>> Users mailing list
>> Users(a)ovirt.org
>>
http://lists.ovirt.org/mailman/listinfo/users
>