KVM: Homogenity of cluster

[Michael Kesper <mk...@schokokeks.org>]

Dear list members,

how important is the homogenity of a kvm cluster?

The docs are very strict here:
- All hosts within a cluster must be homogenous. 
  The CPUs must be of the same type, count, and feature flags.
http://docs.cloudstack.apache.org/en/latest/installguide/hypervisor/kvm.html#system-requirements-for-kvm-hypervisor-hosts

What happens if I integrate different nodes into one cluster as I know
I don't want to do live migration and use them for different workloads?

Would it be possible to tag the differing hosts (e.g. test/prod) so
they don't get mixed or would it really be needed to create a seperate cluster
for that?

N.B.: I also opened a pull request that is influenced by this:
https://github.com/apache/cloudstack-documentation/pull/48
(in short: if all host MUST be homogenous, what sense makes setting CPU features?)

Bye
Michael

Re: KVM: Homogenity of cluster

[Michael Kesper <mk...@schokokeks.org>]

Hi Andrija,

On 25.06.19 12:56, Andrija Panic wrote:
> They need to be if as similar CPU as possible, in order to allow successful
> live migrations - other than that, you could technically combine Intel and
> AMD if you like. Networking bridges also needs to be setup in identical way
> (irrelevant of the possibility different NICs model and speed) - but that
> is zone wide so I guess not even a question.
> 
> In general you want to keep it , CPU wise, as close as possible

I guess load allocation would suffer if CPUs would differ too much in
available load, right?

> - but you
> can have i.e. different generation CPUs (with obviously different CPU
> flags) and then define CPU model to be the lowest common denominator i.e
> define SandyBridge as the model, even though you have both SandyBridge and
> a newer generation CPUs. You DO need to test your CPU models, ACS and
> libvirt (in specific version if libvirt there was some issue with CPU
> flags).

Good hint!
 
> Hope that helps

Indeed, thank you very much!

Michael

Re: KVM: Homogenity of cluster

[Andrija Panic <an...@gmail.com>]

Hi Michael,

They need to be if as similar CPU as possible, in order to allow successful
live migrations - other than that, you could technically combine Intel and
AMD if you like. Networking bridges also needs to be setup in identical way
(irrelevant of the possibility different NICs model and speed) - but that
is zone wide so I guess not even a question.

In general you want to keep it , CPU wise, as close as possible - but you
can have i.e. different generation CPUs (with obviously different CPU
flags) and then define CPU model to be the lowest common denominator i.e
define SandyBridge as the model, even though you have both SandyBridge and
a newer generation CPUs. You DO need to test your CPU models, ACS and
libvirt (in specific version if libvirt there was some issue with CPU
flags).

Hope that helps

Andrija

On Tue, Jun 25, 2019, 09:08 Michael Kesper <mk...@schokokeks.org> wrote:

> Dear list members,
>
> how important is the homogenity of a kvm cluster?
>
> The docs are very strict here:
> - All hosts within a cluster must be homogenous.
>   The CPUs must be of the same type, count, and feature flags.
>
> http://docs.cloudstack.apache.org/en/latest/installguide/hypervisor/kvm.html#system-requirements-for-kvm-hypervisor-hosts
>
> What happens if I integrate different nodes into one cluster as I know
> I don't want to do live migration and use them for different workloads?
>
> Would it be possible to tag the differing hosts (e.g. test/prod) so
> they don't get mixed or would it really be needed to create a seperate
> cluster
> for that?
>
> N.B.: I also opened a pull request that is influenced by this:
> https://github.com/apache/cloudstack-documentation/pull/48
> (in short: if all host MUST be homogenous, what sense makes setting CPU
> features?)
>
> Bye
> Michael
>
>