[pve-devel] Hyperconverged Cloud / Qemu + Ceph on same node

[Stefan Priebe - Profihost AG]

Am 24.07.2018 um 09:25 schrieb Dietmar Maurer:
>> Am 23.07.2018 um 21:04 schrieb Alexandre DERUMIER:
>>> Personally, I think that a vm could take all cpu usage, or memory, and
>>> impact ceph cluster for other vms.
>>>
>>> we should give ceph some kind of (configurable) priority.
>>
>> Yes the problem is or might be. Most users will overcommit the cpu cores
>> with VMs. So theoretically the kvm processes might eat ALL CPU cores.
> 
> Really? I though we have a scheduler that give all processes a change to run.
> 
>> But while running ceph on the same node this should NEVER happen. So my
>> idea is to give cores exclusively to ceph.
> 
> Why exclusively? A scheduler can prioritize specific processes.

I'm not sure how this influences latency - it will move the processes
between cores and still fight against qemu. Even if the scheduler is
fast enough you still have the NUMA problem.

See for example:
http://tracker.ceph.com/projects/ceph/wiki/Tuning_for_All_Flash_Deployments#Ceph-Storage-Node-NUMA-Tuning

Greets,
Stefan