Re: [Qemu-devel] [Qemu-discuss] [problem]-starting a VM causes RT (latency) problem for application running on another VM

[Yan Fridland]

+ qemu-devel

Hello All,

Kindly read this short description of the problem (from the last week) and try 
to assist.
Currently the problem is constantly seen and easily reproduced.

Thank you,
Yan


From: Yan Fridland
Sent: Wednesday, July 27, 2016 5:05 PM
To: 'Yadi'; 'address@hidden'
Subject: RE: [Qemu-discuss] [problem]-starting a VM causes RT (latency) problem 
for application running on another VM


Suppose all the QEMU threads and application of both VMs run on separate CPUs 
(which should be guaranteed by separate vcpus and iothreads pinning in the 
XML), I am trying to think of a resource that can be shared between the 2 VMs. 
The only thing that comes into my mind is memory, that supposed to be isolated 
for each VM, but perhaps its management on the hosts causes such interference.

I saw that the RAM allocation for the VMs is done from the AnonHugePages and 
not from the old mechanism of regular Hugepages. May this somehow relate to my 
problem? What else can it be?

Waiting for your comments,

Thanks,
Yan

From: Yadi [mailto:address@hidden
Sent: Tuesday, July 26, 2016 12:24 PM
To: address@hidden<mailto:address@hidden>; Yan Fridland
Subject: Re: [Qemu-discuss] [problem]-starting a VM causes RT (latency) problem 
for application running on another VM

On 2016年07月26日 15:48, Yan Fridland wrote:
Hi Yadi,

Thanks a lot for your comment.

Actually I am doing emulator pin to the QEMU process in the XML:

<vcpu placement='static'>3</vcpu>
  <cputune>
    <vcpupin vcpu='0' cpuset='1'/>
    <vcpupin vcpu='1' cpuset='2'/>
    <vcpupin vcpu='2' cpuset='3'/>
    <emulatorpin cpuset='8'/>
    <vcpusched vcpus='1-2' scheduler='fifo' priority='99'/>
  </cputune>

BUT, looks like it doesn’t help. Is there any way to make ALL the related QEMU 
threads run within the scope of cores 1,2,3,8 ONLY in the above example?

it is so wired, I have no better ways to handle it since I have no experienced 
the similar issue as your. so, just two suggestion for you:

1: check all qemu thread associated entries under cgroup filesystem,  sure that 
 cpusets are correctly set.
2: if 1 fails, it may means there is nothing left to do, ftrace for schduler.

whatever result, let me know please. I am very interesting in the issue.

Yadi



Thanks,
Yan

From: Yadi [mailto:address@hidden
Sent: Tuesday, July 26, 2016 6:12 AM
To: Yan Fridland; address@hidden<mailto:address@hidden>
Subject: Re: [Qemu-discuss] [problem]-starting a VM causes RT (latency) problem 
for application running on another VM

On 2016年07月25日 23:03, Yan Fridland wrote:

Hello All,



Here is a problem I saw when I ran virsh start to a VM.

I have 2VMs on a server when one of them is running a RT application and the 
other is in shut down state. Then I executed "virsh start" to the second VM and 
saw 2-3 seconds later a big latency problem in my RT application running on the 
first VM.

I want to emphasize that the VMs use dedicated isolated cores of the server and 
I have the relevant configuration on the server and the vcpu pinning section of 
virsh xml.



Can anyone tell me why I experience such behavior and how it can be resolved?



I will highly appreciate your help.



Thank you,

Yan


I am not very sure, but I suggest you to add emulatorpin in xml file, then see 
results. actually,  all qemu associated process can use all processors via 
cgroup configure created by libvirt,

https://libvirt.org/formatdomain.html#elementsCPUTuning