[libvirt] Global domain cpu.cfs_quota_us setup

[Alexander Burluka]

Hi!
Thank you for answer, Daniel, I should excuse for a long delay before 
reply. It was caused by long offline discussion in our company .
Our business logic needs this top level cgroup. Parallels Cloud Server 6 
users have use this
cgroup for a long time and want similar functionality. Also we believe 
that VMWare virtual machines has specified limit:
http://kb.vmware.com/selfservice/microsites/search.do?language=en_US&cmd=displayKC&externalId=1033115
"*Note*: For an SMP virtual machine, the sum of all vCPUs cannot exceed 
the specified limit.
  For example, 4 vCPU virtual machine with a limit of 1200MHz and equal 
load among vCPUs would result in a max of 300MHz per vCPU."
Would anyone mind if we implement this feature as an optional parameter 
in libvirt?


On 12/08/2015 12:27 PM, Daniel P. Berrange wrote:
> On Mon, Dec 07, 2015 at 03:17:53PM +0300, Alexander Burluka wrote:
>> Hello all!
>>
>> I would greatly appreciate your help with some cgroup question. Libvirt xml
>> has cpu tuning section and quota element in it. That parameter sets up
>> cpu.cfs_cgroup_us for every domain's vcpu. I set it to 55555 and has such
>> hierarchy of cgroups:
>> cpu.cfs_quota_us (-1)
>> |
>> |-emulator/cpu.cfs_quota_us (-1)
>> |-vcpu0/cpu.cfs_quota_us (55555)
>> |-vcpu1/cpu.cfs_quota_us (55555)
>> |-vcpu2/cpu.cfs_quota_us (55555)
>> ....
>> \-vcpuN/cpu.cfs_quota_us (55555)
>>
>> Is there any way to set up upper level quota limit via libvirt without
>> setting every vcpu? So I'd like to get next hierarchy:
>> cpu.cfs_quota_us (55555)
>> |
>> |-emulator/cpu.cfs_quota_us (-1)
>> |-vcpu0/cpu.cfs_quota_us (-1)
>> |-vcpu1/cpu.cfs_quota_us (-1)
>> |-vcpu2/cpu.cfs_quota_us (-1)
>> ....
>> \-vcpuN/cpu.cfs_quota_us (-1)
> If you set it at the level up in the hierarchy, then the default case
> ends up with all VMs having the same quota regardless of the number
> of vCPUs they have. By setting it against the vcpu, the quota scales
> as you hotplug/unplug vCPUs
>
>
> Regards,
> Daniel

-- 
Regards,
Alexander Burluka