Lower performance issue on 4CPU AMD server

[flopek]

Hello,

I've got some issue with performance with Win7 x64 Pro as guest system.

I've installed proxmox on HP DL585 server (4x AMD Opteron processors -with NUMA, 64GB of ram, RAID1).

I've created guest VM with 2x 12cores processors (HOST type), 48GB of RAM (statically allocated), VHD as virtio raw device.

I've got serious performance loss, something about 20% in comparission to natively installed W2k8 R2 on the same server.

Maybe you have some tips for me, what I can make to increase performance?

Regards
Krzysiek

 

[spirit]

VHD as virtio raw device.

Don't use VHD in production (it's only to have some compatibility).

convert it to raw file.

(qemu-img convert yourfile.vhd -O raw outputfile.raw)

 

[flopek]

I've make silly mistake

I mean VHD - virtual hard drive. I've used raw of course.

Regards
Krzysiek

 

[spirit]

OK

I've got serious performance loss, something about 20% in comparission

are you talking about cpu, disk or network ?

 

[flopek]

OK



are you talking about cpu, disk or network ?

[/COLOR]

CPU. I've made 3D rendering tests on both VM and host machine. Times differ by about 20%.

I've found tip about assigning cores (http://docs.fedoraproject.org/en-US/Fedora/13/html/Virtualization_Guide/ch25s06.html) in NUMA based systems but when I run virsh list I've got emtpy node list although guests machines are created via web console (up and running).

Any sugestions?

Regards
Krzysiek

 

[spirit]

CPU. I've made 3D rendering tests on both VM and host machine. Times differ by about 20%.

I've found tip about assigning cores (http://docs.fedoraproject.org/en-US/Fedora/13/html/Virtualization_Guide/ch25s06.html) in NUMA based systems but when I run virsh list I've got emtpy node list although guests machines are created via web console (up and running).

Any sugestions?

Regards
Krzysiek

Just to be sure, what vcpu model do you choose for your guest ? (default kvm64 or more advanced ?)

 

[dietmar]

CPU. I've made 3D rendering tests on both VM and host machine. Times differ by about 20%.

Just a guess, but AFAIK 3D rendering usually use hardware acceleration from host, which is not available inside the VM.

 

[flopek]

Just to be sure, what vcpu model do you choose for your guest ? (default kvm64 or more advanced ?)

I've chosen HOST arch.

 

[flopek]

Just a guess, but AFAIK 3D rendering usually use hardware acceleration from host, which is not available inside the VM.

What type of acceleration do you mean?

If I set HOST as guest processor type, isn't it true that guest should support the same features as host machine?

Regards
Krzysiek

 

[spirit]

What type of acceleration do you mean?

If I set HOST as guest processor type, isn't it true that guest should support the same features as host machine?

Regards
Krzysiek

yes, it should be the same.
Maybe it could be memory bandwith slower (or numa miss-placement).

(Just to be sure, you don't use any GPU acceleration (opencl,cuda,....) ?)

 

[dietmar]

what type of acceleration do you mean?

gpu

 

[spirit]

maybe can you try for test the pve 3.10 kernel, I think that new auto numa balacing is include.

(cat /proc/sys/kernel/numa_balancing) = 1

 

[flopek]

gpu

I'm using renderer which is CPU based, so probably it's not the problem.

 

[flopek]

maybe can you try for test the pve 3.10 kernel, I think that new auto numa balacing is include.

(cat /proc/sys/kernel/numa_balancing) = 1

Oh, this sounds interesting!

I will check that option today and give you test results

Thanks
Krzysiek