KVM and cpu limiting

KVM and cpu limiting

[Boris Dolgov]

Hello!

Is it possible to limit cpu usage be VM when using qemu+kvm?

I need one server to run a lot of VMs (much more than cpu cores I
have). Each VM must have it's guaranted cpu time or time slice it must
use.

Is it possible to make such limits? Is it possible to modify
/proc/cpuinfo of VMs?

-- 
Boris Dolgov.

Re: Fwd: KVM and cpu limiting

[Srivatsa Vaddagiri]

> Is it possible to limit cpu usage be VM when using qemu+kvm?

Have you checked cpu controller?

	# mkdir /cpu_control
	# mount -t cgroup -o cpu none /cpu_control

	# cd /cpu_control
	# mkdir vm1
	# mkdir vm2

Then change vm{1,2}/cpu.shares to control how much proportion of CPU capacity
each VM gets. Basically, each vm is (soft-)limtied to:

	its allocated shares / total shares allocated for all VMs.

> I need one server to run a lot of VMs (much more than cpu cores I
> have). Each VM must have it's guaranted cpu time or time slice it must
> use.

This interface is more focused on limiting resource consumption rather than
guaranteeing certain minimum. By carefully controlling how many VMs are run on
the server and how many shares are allocated to each, this can be a good
alternative to provide some "minimum" resources to a VM. Its slightly inflexible
for guaranteeing "minimum" resource to a VM, but that's the best we have now.

> Is it possible to make such limits? Is it possible to modify
> /proc/cpuinfo of VMs?

Why do you want to modify /proc/cpuinfo of VMs?

- vatsa

Re: KVM and cpu limiting

[RW]

> Is it possible to limit cpu usage be VM when using qemu+kvm?
>   

First thing you can do is to only provide one or two
virtual CPUs to a VM. This is normally a not to bad
way to limit CPU usage. But maybe cgroups can help
you more:
http://www.mjmwired.net/kernel/Documentation/cgroups.txt

> I need one server to run a lot of VMs (much more than cpu cores I
> have). Each VM must have it's guaranted cpu time or time slice it must
> use.
>
> Is it possible to make such limits? Is it possible to modify
> /proc/cpuinfo of VMs?
>
>   

You can start Qemu/KVM with different "-cpu" parameter:
bern:~ # qemu-kvm -cpu ?
x86           qemu64
x86           phenom
x86         core2duo
x86            kvm64
x86           qemu32
x86          coreduo
x86              486
x86          pentium
x86         pentium2
x86         pentium3
x86           athlon
x86             n270

Or just provied "-cpu host". Then you get the same output
inside the guest as the host.

- Robert

Re: Fwd: KVM and cpu limiting

[Boris Dolgov]

On Fri, Jul 2, 2010 at 11:57 AM, Srivatsa Vaddagiri <vatsa@in.ibm.com> wrote:
>> Is it possible to limit cpu usage be VM when using qemu+kvm?
>
> Have you checked cpu controller?
It is very interesting. Looks like it is something, that I need.
So, limiting qemu process will limit cpu usage for the container?
I think, i will be able to configure it to provide guaranted cpu to
containers in that way.

But what happens with the VM when qemu process is paused by the
scheduler? Will this time be counted as idle time for VM's cpu, or it
will be added to the state, which was used when qemu process was
paused?

> Why do you want to modify /proc/cpuinfo of VMs?
Just for information about the limits for the VMs' owners.

-- 
Boris Dolgov.

Re: Fwd: KVM and cpu limiting

[Srivatsa Vaddagiri]

On Fri, Jul 02, 2010 at 08:38:37PM +0400, Boris Dolgov wrote:
> On Fri, Jul 2, 2010 at 11:57 AM, Srivatsa Vaddagiri <vatsa@in.ibm.com> wrote:
> >> Is it possible to limit cpu usage be VM when using qemu+kvm?
> >
> > Have you checked cpu controller?
> It is very interesting. Looks like it is something, that I need.
> So, limiting qemu process will limit cpu usage for the container?
> I think, i will be able to configure it to provide guaranted cpu to
> containers in that way.

Note that shares control how much cpu bandwidth a VM gets _relative_ to other
active VMs and hence is not a absolute measure of how much CPU capacity is
provided to a particular VM. If the number of VMs deployed on a server 
and their relative priority remains static in general, then shares can be 
used as both "guarantee" and "limit" mechanism.

> But what happens with the VM when qemu process is paused by the
> scheduler? Will this time be counted as idle time for VM's cpu,

No - it should be accounted as "runqueue wait time".

> or it will be added to the state, which was used when qemu process was
> paused?

The runqueue wait time should reflect as "steal" time for the VMs. top command
should show the steal time for a VM (look for 'st'), although I am not 100% sure
whether steal time is reported in case of KVM guests.

> > Why do you want to modify /proc/cpuinfo of VMs?
> Just for information about the limits for the VMs' owners.

Sounds like a cool idea - although not sure whether cpuinfo is the right place.
You can try sending a patch and find out what the community thinks of it!

- vatsa