Intel Xeon E-2176G - Performance problems and wrong MHz measuring

[sharkIT]

Hello,

we have some performance problems with a test enviroment. The enviroment is a Xeon E-2176G server with vSphere 6.7U2.

Initial the test was about NVMe performance. But we didnt get good results so we did some mearuements and found out the the CPU load average is really high.

We have four Server 2016 VM`s for the performance tests. It seems that just little MS OS systemload is not easy to handle for the E-2176G.

The really strage thing and the main question of this post is, if we produce CPU load in the VM`s and we get a host esxtop CPU load average of >1 the vSphere GUI shows a the same time a host CPU load of about 20%.

It seems that the host is using just 1/5th of the available CPU MHz performance. At same time esxtop shows a core util of every core with 100%.

Every setting is performance optimized (power, bios...). Whats wrong? I have no more ideas and think about something like

- P-state, C-state problems

Do you have experiance with XEON E-21XX series CPU`s. Up to now we just used E-26XX oder Xeon Silver series. It seems that the CPU cant provide 3,7 GHz on every (6) core.

Thank you for your feedback.
Best Regards

ITKOA

[blazilla]

Hi,

try to disable any CPU power saver settingd, e.g. go to the BIOS/UEFI and set it to something like „High Performance“.

[sharkIT]

Hi Blazilla,

the BIOS settings are High performance, but as well I think there is some problem with CPU power options. I tried default settings, performance settings (Fujitsu performance guide) both with same results. The vSphere host power options are also high performance.

[blazilla]

Did you set the power settings in Windows to „High Performance„?

[sharkIT]

As well in Windows. We tried different Custom Images and original vSphere Images. We can produce a esxtop load of 100% but 100% esxtop load seems to be just 20% of the available MHz ressources. So if its a engergy problem it has something to do beteween the hardware and the hypervisor. I dont think its a problem between VM and hypervisor because esxtop shows a 100% CPU load. It seems that esxtop has a differnt measurement procedure compared to the GUI.

[blazilla]

6 pCores and 6 Threads are fully utilized. That‘s what I would expect. can you provide more detailed output from the CPU screen of esxtop.

[sharkIT]

Right, so esxtop show full utilized cores and threads.  At same time the consumed MHz of the vSphere Host CPU is just 21% (right side of the screenshot).

If I do a stress test I can utilize the system to 100% in esxtop but just 22% in the vSphere GUI host CPU measurement.

For me it seems like the host is not using the whole MHz frequenz or dont use every core. Really strange are two parameters

1. 100% esxtop are 22% in the vSphere GUI (4,7GHz)

2. 100% esxtop core util shows a same time a maximium of 10% pcpu utli %

Best Regards

[blazilla]

With how many VMs do you test? How is every VM sized and what tool do you use?

[sharkIT]

There are 4 MS OS 2016 VM`s on the system. The idle load is <5%.

I did tests with CPU benchmark tools like cpu-z.

First test 4 VM`s, 2 cores per VM, CPU stress test on every vm running

Second test 1 VM with 12 cores, CPU stress test just on one vm

I always get the same results/limits

- max 4700MHz from availabe ~22000MHz in vSphere GUI

- max PCPU used % 10 in esxtop

at the same time the core util % is about 100% and the CPU load average >2 so super overloaded

Best Regards

ITKOA

[blazilla]

In both cases you used more vCPUs then pCores we’re available. What about six VMs with one vCPU each?