[Xenomai-help] Does qemu-kvm impact realtime performance?

[Xenomai-help] Does qemu-kvm impact realtime performance?

[Henri Roosen]

Hi all,

Does anyone know if running qemu-kvm impacts the realtime Xenomai
performance of the host system?

I have a multicore x86 target that runs a Xenomai enabled kernel and a
realtime application. Next to the realtime application I would like to
run qemu-kvm (on one of the cores) that runs some non-realtime guest
os. So I would like to know if qemu-kvm have impact on the
latencies/jitter of the realtime threads?

Thanks,
Henri

Re: [Xenomai-help] Does qemu-kvm impact realtime performance?

[Gilles Chanteperdrix]

On 05/10/2012 04:26 PM, Henri Roosen wrote:
> Hi all,
> 
> Does anyone know if running qemu-kvm impacts the realtime Xenomai
> performance of the host system?
> 
> I have a multicore x86 target that runs a Xenomai enabled kernel and a
> realtime application. Next to the realtime application I would like to
> run qemu-kvm (on one of the cores) that runs some non-realtime guest
> os. So I would like to know if qemu-kvm have impact on the
> latencies/jitter of the realtime threads?

I would say it should have no direct impact, we should not expect kvm to
create hard irqs masking sections. However, as usual, running a whole OS
and applications inside another OS probably has a huge impact on cache,
and so should cause more jitter than running xenomai on an idle system.

-- 
					    Gilles.

Re: [Xenomai] Does qemu-kvm impact realtime performance?

[Jan Kiszka]

On 2012-05-11 05:22, Gilles Chanteperdrix wrote:
> On 05/10/2012 04:26 PM, Henri Roosen wrote:
>> Hi all,
>>
>> Does anyone know if running qemu-kvm impacts the realtime Xenomai
>> performance of the host system?
>>
>> I have a multicore x86 target that runs a Xenomai enabled kernel and a
>> realtime application. Next to the realtime application I would like to
>> run qemu-kvm (on one of the cores) that runs some non-realtime guest
>> os. So I would like to know if qemu-kvm have impact on the
>> latencies/jitter of the realtime threads?
> 
> I would say it should have no direct impact, we should not expect kvm to
> create hard irqs masking sections. However, as usual, running a whole OS
> and applications inside another OS probably has a huge impact on cache,
> and so should cause more jitter than running xenomai on an idle system.

On a modern VT-x or AMD-V system, the impact is not different from
running some cache hungry application on the host besides a Xenomai
load. Recent processors can track which cache entry belongs to which
memory mapping and optimize the eviction regarding this. But if your RT
frequency is sufficiently low and you share the caches, you will still
see impact - with or without virtualization.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux

Re: [Xenomai] Does qemu-kvm impact realtime performance?

[Henri Roosen]

On Mon, May 21, 2012 at 1:35 PM, Jan Kiszka <jan.kiszka@siemens.com> wrote:
> On 2012-05-11 05:22, Gilles Chanteperdrix wrote:
>> On 05/10/2012 04:26 PM, Henri Roosen wrote:
>>> Hi all,
>>>
>>> Does anyone know if running qemu-kvm impacts the realtime Xenomai
>>> performance of the host system?
>>>
>>> I have a multicore x86 target that runs a Xenomai enabled kernel and a
>>> realtime application. Next to the realtime application I would like to
>>> run qemu-kvm (on one of the cores) that runs some non-realtime guest
>>> os. So I would like to know if qemu-kvm have impact on the
>>> latencies/jitter of the realtime threads?
>>
>> I would say it should have no direct impact, we should not expect kvm to
>> create hard irqs masking sections. However, as usual, running a whole OS
>> and applications inside another OS probably has a huge impact on cache,
>> and so should cause more jitter than running xenomai on an idle system.
>
> On a modern VT-x or AMD-V system, the impact is not different from
> running some cache hungry application on the host besides a Xenomai
> load. Recent processors can track which cache entry belongs to which
> memory mapping and optimize the eviction regarding this. But if your RT
> frequency is sufficiently low and you share the caches, you will still
> see impact - with or without virtualization.
>
Thanks Jan and Gilles for your replies!
We will do some testing and measurements on this in a later project
and I will post some results then.

Thanks,
Henri.
> Jan
>
> --
> Siemens AG, Corporate Technology, CT T DE IT 1
> Corporate Competence Center Embedded Linux