CPU usage for credit2

CPU usage for credit2

[Yuehai Xu]

Hi,

Suppose 2 VMs both of which just has a single VCPU share a single
PCPU, if I run CPU intensive program in both of the VMs, the
percentage of CPU usage from xentop is almost 100% for both of the
VMs. I got confusing because since 2 VCPU shares a single PCPU, the
percentage for each one should be around 50%, which is the result if I
change the scheduler from credit2 to credit.

Does it because credit2 change something or is this a bug for xentop or credit2?

Thanks,
Yuehai

Re: CPU usage for credit2

[George Dunlap]

I assume that your system has more than 1 PCPU, and that you're using 
pinning to try to limit the VM to 1 PCPU?

At the moment, the credit2 scheduler doesn't implement pinning -- cpu 
affinity is entirely ignored.  (It is an experimental scheduler, after 
all.)  I think theoretically making pinning function shouldn't be too 
hard; but making the credit algorithm work with an arbitrary combination 
of affinities would require some careful thinking, and might not even be 
possible.

If you want, you can try to use cpu pools to isolate the two VMs to a 
single PCPU.  I haven't tried credit2 with pools yet, but in theory it 
should work.

  -George

On 23/09/10 14:48, Yuehai Xu wrote:
> Hi,
>
> Suppose 2 VMs both of which just has a single VCPU share a single
> PCPU, if I run CPU intensive program in both of the VMs, the
> percentage of CPU usage from xentop is almost 100% for both of the
> VMs. I got confusing because since 2 VCPU shares a single PCPU, the
> percentage for each one should be around 50%, which is the result if I
> change the scheduler from credit2 to credit.
>
> Does it because credit2 change something or is this a bug for xentop or credit2?
>
> Thanks,
> Yuehai

Re: CPU usage for credit2

[Yuehai Xu]

On Fri, Sep 24, 2010 at 7:38 AM, George Dunlap
<George.Dunlap@eu.citrix.com> wrote:
> I assume that your system has more than 1 PCPU, and that you're using
> pinning to try to limit the VM to 1 PCPU?

Yes, you are correct. In order to test the performance of credit2 for
the problem of mix work load, I pin VCPUs from VMs to a PCPU while the
VCPU of Dom0 pins to another PCPU.

>
> At the moment, the credit2 scheduler doesn't implement pinning -- cpu
> affinity is entirely ignored.  (It is an experimental scheduler, after all.)
>  I think theoretically making pinning function shouldn't be too hard; but
> making the credit algorithm work with an arbitrary combination of affinities
> would require some careful thinking, and might not even be possible.
>

So, the result that I got, which shows almost 100% usage of CPU for
each VCPU, is actually because the VCPUs use different PCPUs instead
of a single one, is that correct?


> If you want, you can try to use cpu pools to isolate the two VMs to a single
> PCPU.  I haven't tried credit2 with pools yet, but in theory it should work.

Ok, I will try it and tell you the result once I get it.

Thanks,
Yuehai

>
>  -George
>
> On 23/09/10 14:48, Yuehai Xu wrote:
>>
>> Hi,
>>
>> Suppose 2 VMs both of which just has a single VCPU share a single
>> PCPU, if I run CPU intensive program in both of the VMs, the
>> percentage of CPU usage from xentop is almost 100% for both of the
>> VMs. I got confusing because since 2 VCPU shares a single PCPU, the
>> percentage for each one should be around 50%, which is the result if I
>> change the scheduler from credit2 to credit.
>>
>> Does it because credit2 change something or is this a bug for xentop or
>> credit2?
>>
>> Thanks,
>> Yuehai
>
>

Re: Re: CPU usage for credit2

[George Dunlap]

On Fri, Sep 24, 2010 at 3:14 PM, Yuehai Xu <yuehaixu@gmail.com> wrote:
> So, the result that I got, which shows almost 100% usage of CPU for
> each VCPU, is actually because the VCPUs use different PCPUs instead
> of a single one, is that correct?

Yes, that's right.

>> If you want, you can try to use cpu pools to isolate the two VMs to a single
>> PCPU.  I haven't tried credit2 with pools yet, but in theory it should work.
>
> Ok, I will try it and tell you the result once I get it.

Cool -- good luck!

 -George