[PVE-User] live migration amd - intel

[mj]

Hi Stefan,

Thanks for your reply!

MJ

On 2/26/20 4:14 PM, Stefan Reiter wrote:
> Hi!
> 
> On 2/26/20 3:35 PM, mj wrote:
>> Hi,
>>
>> Just to make sure I understand something. We have an identical 
>> three-node hyperconverged pve cluster, on Intel Xeon CPU's.
>>
>> Now we would like to expand, and we are investigating what path to 
>> choose.
>>
>> The way we understand it is: if we would like to do live migrations 
>> between pve hosts, the servers need to have similar CPUs, or otherwise 
>> we need to virtualise the CPU as well. (so set CPU type to kvm64)
>>
> 
> Note that even kvm64 does not really work for cross-vendor 
> live-migrations (i.e. AMD <-> Intel as mentioned in subject). This was 
> recently discussed on the pve-devel list [0], as well as an older bug 
> report [1].
> 
> [0] https://pve.proxmox.com/pipermail/pve-devel/2020-February/041750.html
> [1] https://bugzilla.proxmox.com/show_bug.cgi?id=1660
> 
>> In case it matters: 90% of our VMs are debian 9/10 hosts.
>>
>> While we could do change to kvm64, we currently use cpu type 'host', 
>> and we wonder how much performance kvm64 would cost us, and what 
>> potential other drawbacks this could have.
>>
> 
> Aside from the above, the performance impact of switching to kvm64 is 
> largely dependent on the software you are running within your VM. Raw 
> CPU performance will stay the same, but most hardware acceleration (e.g. 
> AES for encryption or AVX) will not be available to the guest.
> 
> If the CPUs are from the same vendor, you can most likely enable the 
> older generation CPU type for your VMs and still get away with 
> live-migration. This way you can use _most_ acceleration features, and 
> are just missing out on any new ones introduced in the newer CPU gen.
> 
>> Has anyone ever done much testing on this subject? Anyone with 
>> interesting insights / knowledge / experiences to share on this subject?
>>
>> MJ
> 
> Hope that helps,
> 
> Stefan
>