Test vCPU configuration to get best performance out of a VM. What configuration shall I use. #vCPU @vExpert @VMware #virtualmachine @FrankDenneman #numa @mwVme

We start with what our goals are

We have a vm that want to take as much cpu power as possible from the host.

But how shall we configure the VM to make the most of it. You have read about numa nodes and what that means? right? If not look, at the bottom of this post you will find find some nice links.

If you really want to deep dive regarding this read anything that Frank Denneman as written, great stuff!

Also take a look at https://notesfrommwhite.net/ from @mwVme great stuff and do great summary of VMware stuff.

Questions

Shall I use vCPU and make it equal to Logical Processors?
Shall I use vCPU and make it equal to Physical socket?
Whats is sockets on a VM?
How many cores per socket on the VM?
So many questions, Lets try the different configurations and see!

I have to answer a few questions during the tests

How many vCPU does the VM see?
How many numa nodes does the VM see?
What is the performance on the VM?

Lets begin

We start with the stuff that I can not change.

We have a server with 2 physical CPU:s
12 Cores per CPU in total 24 cores
Then we have hyper threading enabled so we get 48 Logical processors.

Host configuration:

So lets see what are the different configuration that we wanna try out for this VM.

The Windows server is 2019 with 8GB of memory. The picture says 4 but I change that after the screenshot.

In the test vm we have 8 GB ram 40 GB. Newly installed Windows 2019 server. The pictures are related to the the test with 48 vCPU and 48 Cores per socket.

Server config

So first of whats does Windows see?

And how many numa nodes can Windows see?

Then we run CPU-Z just to verify:

Now to performance. We use passmarks program to test perfromance.

I don’t wanna bore you out with a lot of picture so you have to believe the excel sheet.

The winner in my case was 48 vCPU and 24 Cores. That will be 2 sockets on the VM and 2 NUMA nodes.

vCPU	Cores	Sockets	Virtual proc. in VM	Sockets in vm	Numa node in vm	CPU Mark	Integer Math	Floating Point	Prime Nr	SSE	Compression	Encryption	Physiscs	Sorting	Single Thread	Cross-plattform	Total
48	1	48	48	48	2	21358	99308	44627	107	24308	404217	6522	1244	56132	1652	52738	712213
48	24	2	48	2	2	22922	92605	50481	107	28282	431481	8166	1098	57940	1652	53816	748550
48	48	1	48	1	2	17121	103058	53062	103	9509	250094	6835	1233	55052	1668	54289	552024
24	1	24	24	24	2	21734	79003	58336	186	28258	365648	5979	1823	51786	1678	61231	675662
24	12	2	24	2	2	21798	78712	57328	173	28341	364276	6147	1802	51361	1685	59729	671352
24	24	1	24	1	2	21788	79391	57852	179	29129	361213	6116	1831	51918	1645	60664	671726
12	1	12	12	12	1	12048	40685	27247	95	14651	182278	2987	1256	25113	1669	30541	338570
12	6	2	12	2	1	12720	41196	30045	104	15852	196008	3134	1301	27784	1673	32882	362699
12	12	1	12	1	1	12705	40564	29850	103	15553	193320	3214	1263	27248	1671	32222	357713
20	2	10	20	10	2	19596	68056	49869	182	25632	318011	5240	1875	45402	1674	54491	590028

Did not get the table nice in WordPress, you can not be good at all the things. So here is a screen shot.

Some sizing rules from vmware blog

Picture below is sizing av VM if the VM has less memory than one of the sockets.

Picture below is sizing av VM if the VM has more memory than one of the sockets.

Conclusions:

If you have a single VM on host, the best configuration is to match the physical setup. In my case was

48 vCPU and 24 Cores per socket

I did not look at the application side, that can be my next project if i get the feeling.

I also only used passmarks application to do the test. As you can see the best result in this test was 48 vCPU 24 Cores and results in 2 sockets. So it matches the physical hardware.

Some numbers are better on different configurations, but I have only looked at total numbers. If you want to calculate prime number the best configuration would be 24 vCPU 1 core for the VM.

And over allocation is also something to think about when you have multiple vm:s on same host.

If you enable Hotadd on a VM numa is disabled.

Regarding the sockets, it is purely based on how the CPUs are presented to the guest.  Whether you choose dual sockets single core each, or single socket with dual core, vSphere treats those as the same, you get 2 cores of CPU time.

Links:

https://frankdenneman.nl/2016/12/12/decoupling-cores-per-socket-virtual-numa-topology-vsphere-6-5/

https://www.passmark.com/products/performancetest/download.php

https://blogs.vmware.com/performance/2017/03/virtual-machine-vcpu-and-vnuma-rightsizing-rules-of-thumb.html

CPU Benchmarks value:

https://www.cpubenchmark.net/cpu_test_info.html

Thank you for reading this far,

Keep hacking!

//Roger