Re: [Qemu-discuss] Latency degradation on VM due to several QEMU processes on the host.

[Yan Fridland]

Hi Alex,

Thanks a lot for the information below.

It really may improve some RT capabilities of my applications running on VMs 
and honestly I am doing 90% from what is mentioned there.

But, one of the problems I saw, to make long story short, if you want low 
latencies, you have to isolate all the cores and run each VCPU on a dedicated 
core and each VCPU thread on different core and each QEMU on different core. 
This way I will run out of cores really fast :), which is extremely problematic 
for me. For example in the ppt I saw:

"Run VCPU threads on dedicated CPUs"
"Run KVM guest VCPUs on isolated CPUs"

If I have 3 vcpus this means that I will need 6 dedicated cores on the host ?? 

Anyway, I will try running QEMUs with "mlock=on" and do some configuration 
updates e.g. idle=poll and not mwait as I have now, but my main questions 
remained:
        1. How can I run several QEMUs on a single core and still keep low 
latencies on the VMs ? Is "mlock=on" the solution for this?
        2. QEMU generates different threads running on different cores (besides 
VCPU threads running on the affinity cores, defined in a pinning section of the 
virsh XML). May this harm latency as well? May and should I do some special 
affinity to them or change priority, etc ... ?


Sincerely,
Yan

-----Original Message-----
From: Qemu-discuss [mailto:address@hidden On Behalf Of Alex
Sent: Sunday, June 19, 2016 3:40 PM
To: address@hidden
Subject: Re: [Qemu-discuss] Latency degradation on VM due to several QEMU 
processes on the host.

Hi,
Take a look at Qemu's realtime feature, might be what you need.

https://www.redhat.com/archives/vfio-users/2015-September/msg00143.html

http://events.linuxfoundation.org/sites/events/files/slides/kvmforum2015-realtimekvm.pdf

/--Regards, Alex/

On 19/06/16 14:38, Yan Fridland wrote:
> Hi,
>
> Can you please comment on the overhead QEMU brings in terms of latency 
> and CPU utilization?
>
>
>
> I am running a highly latency sensitive application on my VMs and I 
> see some degradation in performance when running 2 QEMU processes on a 
> single core, or even when I run 1 QEMU process on a single core 
> without a dedicated affinity and RT priority to the QEMU process.
>
>
>
> Can I make the QEMU lighter to be able to run 2,3 or more QEMUs on a 
> single core? Is it matter of XML configuration / affinity / priority 
> or anything else?
>
>
>
> Thank you,
>
> Yan
>