CPU performance comparison Win10-Ubuntu

[ogghi]

Hey Guys,

we are facing a strange issue of performance difference between a Ubuntu Server 16.04.1 VM and a Win10 Pro VM.

I used geekbench4 on both machines to test.

Hardware configurations and scores:

Client win10, qemu cpu, no-numa, 2s20c, 16gb-no-ballooning -> 8200 multicore
Client ubuntu, qemu cpu, no-numa, 2s20c, 16gb-no-ballooning -> 18882 multicore

Any ideas how this could come?

 

[udo]

Hi,
how many CPUs/Cores has your host?
Why do you disabled Numa?

Udo

 

[Ashley]

Have you checked to make sure all your cores show in Windows OS?

 

[ogghi]

Hi, thanks for the replies.

I tried also with NUMA, not much of a change.

Alls 40 cores I'm passing through are showed in Win and Ubuntu.

The host hast 4x Xeon E7-4850 with 10c/20threads

 

[Ashley]

Do you have the CPU model set as host pass through?

 

[ogghi]

I tried also with CPU setting "host" instead of "Default (kvm64)" but there's no reasonable difference.

 

[Ashley]

How does task manager CPU graphs look during benchmark?

What is Windows 10 power profile set to, do remember only server editions of Windows are designed for use within a VM so may never get the same performance.

But there will be I'm sure some Optimisations to get an increase. However right now just could be the benchmark within Windows not making use of all threads and not a "Proxmox" issue.

 

[ogghi]

Hi, tried the same with Windows10 and "High performance" profile, no change.
Some other tests:
Client win2012R2, qemu cpu, no-numa, 2s20c, 16gb-no-ballooning -> 10000 multicore
Client win7, qemu cpu, no-numa, 2s20c, 16gb-no-ballooning -> 9991 multicore

Weird or normal?

Forgot to add: In task manager it shows all cores utilized!

 

[ogghi]

Could anybody else do a simple comparison between Windows and Linux to see if it's an issue on my side, or just "overhead"?

 

[ogghi]

Is there any additional test that I could do?

 

[ogghi]

Nobody with an idea? I have to convince my boss to buy proxmox, but right now I can't...

 

[dcsapak]

I tested here with a smaller config (only 1 Socket 2 cores , CPU on "host" and 4GB Memory, everything else default), and had not so big of a difference

Windows 10 Pro:

Single Core: 5170
Multi Core: 8868

Ubuntu 16.04:

Single Core: 5456
Multi Core: 9982

so linux is a bit faster but not much,
the hardware here is a Intel Core i7-6700k @ Stock speeds
with 32GB Memory

 

[ogghi]

Interesting, the big difference is just in the multicore performance!
https://browser.geekbench.com/v4/cpu/1808549
https://browser.geekbench.com/v4/cpu/1810691

Maybe I could tune CPU host settings, NUMA or something similar?

Otherwise we might actually need to look for a different/additional server to build under windows...

 

[ogghi]

Ok, did some more testing:

I installed esxi on the same host and configured an ubuntu and a win2012r2 VM:
https://browser.geekbench.com/v4/cpu/2047977
https://browser.geekbench.com/v4/cpu/2048634

So either geekbench is questionable as a comparison or windows is just much worse in performance.

If I find the time I might go ahead and install physically Win and ubuntu for half a day.

 

[ogghi]

Hi folks,
it got a bit better by disabling hyperthreading!
Win2012 with 4s10c -> 11810 https://browser.geekbench.com/v4/cpu/2490562

Win10 with 2s10c -> 8867 https://browser.geekbench.com/v4/cpu/2490373

Ubuntu with 4s 10c -> 28959 https://browser.geekbench.com/v4/cpu/2490139

 

[dietmar]

Ubuntu with 4s 10c -> 28959 https://browser.geekbench.com/v4/cpu/2490139

Do you really have 40 cores? It usually slows down things if you assign more cores than you have (to a single VM).

 

[ogghi]

Yep, there are 4CPUs with 10 cores each.
Ubuntu is actually fine here, it's windows through virtualization that is suuuper slow...

 

[JohnD]

The windows scheduler is just unable to handle so many cores correctly. It is the same on physical servers and has nothing to do with proxmox.

 

[zima]

on r910 4xX7560 with about 20% cpus used by other vm's
https://browser.geekbench.com/v4/cpu/2611245

on other r910 without any running vm's
win - https://browser.geekbench.com/v4/cpu/2611348

ubuntu - https://browser.geekbench.com/v4/cpu/2611539
bare-metal debian (proxmox) https://browser.geekbench.com/v4/cpu/2611723
kvm ~90% of bare-metal performance, unfortunately i don't have windows on bare-metal right now to compare windows score.

                           node0           node1           node2           node3
numa_hit               280149326       290805320       267824125       226700732
numa_miss                      0               0               0               0
numa_foreign                   0               0               0               0
interleave_hit             18041           18440           18344           18449
local_node             280142722       290784647       267804011       226680404
other_node                  6604           20673           20114           20328

A year ago i have done some test with win vm with 32cores/64threads and windows slows down when cores counts goes up.
In cinebench single cpu performance is virtually the same, but multi-core ->

2578 bare-metal vs 2035 kvm ~80% of bare-metal performance

 

[ogghi]

The windows scheduler is just unable to handle so many cores correctly. It is the same on physical servers and has nothing to do with proxmox.

Wow, that's sad then, but good to know!
Next build server hardware: Less cores, more clock speed!