[libvirt] [PATCH 0/7] Per domain bandwidth settings

[Alexander Burluka]

Oh, I got it, your example is really problem case, but it's a cpuset 
specific.
cfs_period_us and cfs_quota_us are a little different, they limit only 
bandwidth
and won't cause anything similar.
For example:
A->B
A cpu.cfs_period_us is set to 100000
If you will try to write 200000 to B cpu.cfs_period_us, you will get a 
write error.
(I've tried this on 3.10 Linux kernel). So child period and quota cannot be
greater than a parent ones (of course if parent constraints are set they 
are ignored).
The main reason to do this knob is flexibility and behavior similar to 
VMWare and
our Parallels Cloud Server 6. Our customers really want it.
By the way, period and quota can be set on running domain.

On 01/13/2016 01:42 PM, Michal Privoznik wrote:
> On 13.01.2016 11:17, Alexander Burluka wrote:
>> You are absolutely right, this patchset allows to set domain top-level
>> cpu.cfs_period_us and cpu.cfs_quota_us cgroups. Can you please
>> explain problem case a little bit more detailed? This code does not
>> affect top-level cpuset cgroup, only quota and period, so there is no
>> visible troubles for me. Thank you!
> Well, I would have to dig into the patches, but in theory, if you have
> the following cgroup layout:
>
> A->B->C
>
> where A is the top-level cgroup, B is child of A, C is child of B, then
> B is inherently restricted by A, C is inherently restricted by B (and
> transitively by A too). Therefore if for instance cpuset in A is set to
> 0-3, B can be only as good as A or more restrictive. So values for
> cpuset in B must be a subset of those in A. And so on. The problem we
> were facing just recently and that I'm mentioning was, that in the
> picture, libvirt puts vCPUs into C cgroup and don't touch A or B. Now
> when user wants to pin vCPUs onto different cpuset, we can simply just
> change it in C cgroup. If we, however, were to set cpuset in B too, it
> would be impossible for us to change C due to reasoning above.
>
> Now, I am not familiar with CFS and we probably don't allow tuning it at
> runtime either. I just want to make sure the reason I am mentioning
> above is kept in picture when touching our CGroups code.
>
> If I'm completely off, and just blabbing off the lines, disregard me. I
> will probably learn more once I'm through the patches.
>
> Michal

-- 
Regards,
Alexander Burluka