vCPU scalability for linux VMs

vCPU scalability for linux VMs

[Alec Istomin]

Gentlemen,
 Reaching out with a non-development question, sorry if it's not
 appropriate here.
 
 I'm looking for a way to improve Linux SMP VMs performance under KVM.

 My preliminary results show that single vCPU Linux VMs perform up to 10
 times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
 8 core pre-Nehalem server). I suspect that I'm missing something major
 and look for any means that can help improve SMP VMs performance.


VMs are started using:
qemu-kvm -m $ram -smp $cpus -name $name -drive file=${newimg},boot=on,cache=writethrough  -net nic,macaddr=$mac,vlan=0,model=virtio -net tap,script=/kvm/qemu-ifup,vlan=0,ifname=kvmnet$i -parallel none -usb -k en-us  -monitor pty -serial pty -nographic -daemonize -snapshot


KVM Host Environment (redhat 5 based):
# uname -r
2.6.18-194.el5
# rpm -qa|grep kvm
kvm-83-164.el5
kvm-tools-83-164.el5
kmod-kvm-83-164.el5
kvm-qemu-img-83-164.el5



Thank you,
 Alec

Re: vCPU scalability for linux VMs

[Avi Kivity]

On 05/05/2010 04:45 AM, Alec Istomin wrote:
> Gentlemen,
>   Reaching out with a non-development question, sorry if it's not
>   appropriate here.
>
>   I'm looking for a way to improve Linux SMP VMs performance under KVM.
>
>   My preliminary results show that single vCPU Linux VMs perform up to 10
>   times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
>   8 core pre-Nehalem server). I suspect that I'm missing something major
>   and look for any means that can help improve SMP VMs performance.
>
>
>    

So you have a total of 32 vcpus on 8 cores?  This is known to be 
problematic.  You may see some improvement by enabling hyperthreading.

There is ongoing work to improve this.

-- 
error compiling committee.c: too many arguments to function

Re: vCPU scalability for linux VMs

[Alec Istomin]

On Wednesday, May 5, 2010 at 04:43:32 -0400, Avi Kivity wrote:

> So you have a total of 32 vcpus on 8 cores?  This is known to be 
> problematic.  You may see some improvement by enabling hyperthreading.

exactly, 32 vCPUs on 8 core hardware that doesn't support hyperthreading
(Clovertown E5335)

> There is ongoing work to improve this.

If interested - let me know when is it going to be a good time/build to
do the regression on this.

Re: Fwd: vCPU scalability for linux VMs

[Srivatsa Vaddagiri]

> Gentlemen,
>  Reaching out with a non-development question, sorry if it's not
>  appropriate here.
> 
>  I'm looking for a way to improve Linux SMP VMs performance under KVM.
> 
>  My preliminary results show that single vCPU Linux VMs perform up to 10
>  times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
>  8 core pre-Nehalem server). I suspect that I'm missing something major
>  and look for any means that can help improve SMP VMs performance.

Could you provide more details of what workload was used to compare this
performance?

- vatsa

Re: vCPU scalability for linux VMs

[Alec Istomin]

On Wednesday, May 5, 2010 at 13:27:39 -0400, Srivatsa Vaddagiri wrote:
>>  My preliminary results show that single vCPU Linux VMs perform up to 10
>>  times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
>>  8 core pre-Nehalem server). I suspect that I'm missing something major
>>  and look for any means that can help improve SMP VMs performance.

> Could you provide more details of what workload was used to compare this
> performance?

It's Dell's DVD-Store LAMP stack test suite with Medium database (about
1 GB MySQL data) . 8x 2Gb VMs (on 16GB RAM host) with different vCPU
configuration simultaneously loaded by external web clients.

Regards,
 Alec

Re: vCPU scalability for linux VMs

[Srivatsa Vaddagiri]

On Wed, May 05, 2010 at 12:31:11PM -0700, Alec Istomin wrote:
> 
> On Wednesday, May 5, 2010 at 13:27:39 -0400, Srivatsa Vaddagiri wrote:
> >>  My preliminary results show that single vCPU Linux VMs perform up to 10
> >>  times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
> >>  8 core pre-Nehalem server). I suspect that I'm missing something major
> >>  and look for any means that can help improve SMP VMs performance.
> 
> > Could you provide more details of what workload was used to compare this
> > performance?
> 
> It's Dell's DVD-Store LAMP stack test suite with Medium database (about
> 1 GB MySQL data) . 8x 2Gb VMs (on 16GB RAM host) with different vCPU
> configuration simultaneously loaded by external web clients.

Ok ..Do you have any more data on how the vCPU scalability looks like (how
the throughput scales with increasing vCPUs)?

- vatsa

Re: vCPU scalability for linux VMs

[Alec Istomin]

[-- Attachment #1: Type: text/plain, Size: 1226 bytes --]

On Wednesday, May 5, 2010 at 01:53:55 -0400, Srivatsa Vaddagiri wrote:
> On Wed, May 05, 2010 at 12:31:11PM -0700, Alec Istomin wrote:
>> 
>> On Wednesday, May 5, 2010 at 13:27:39 -0400, Srivatsa Vaddagiri wrote:
>> >>  My preliminary results show that single vCPU Linux VMs perform up to 10
>> >>  times better than 4vCPU Linux VMs (consolidated performance of 8 VMs on
>> >>  8 core pre-Nehalem server). I suspect that I'm missing something major
>> >>  and look for any means that can help improve SMP VMs performance.
>> 
>> > Could you provide more details of what workload was used to compare this
>> > performance?
>> 
>> It's Dell's DVD-Store LAMP stack test suite with Medium database (about
>> 1 GB MySQL data) . 8x 2Gb VMs (on 16GB RAM host) with different vCPU
>> configuration simultaneously loaded by external web clients.

> Ok ..Do you have any more data on how the vCPU scalability looks like (how
> the throughput scales with increasing vCPUs)?

See attached for 1 vs 2 vs 4 vCPUs consolidated throughput from multiple
VMs with the same client payload. Selected DVD-Store (LAMP) scalability
is mixed into here too and I wonder what type of test would you
recommend to receive cleaner results?

Regards,
 Alec

[-- Attachment #2: kvm-vcpu-graph.png --]
[-- Type: IMAGE/PNG, Size: 18225 bytes --]