After debugging a recent issue with using VMware Workstation and Intel Hybrid CPUs, it gave me an idea about an experiment to try with ESXi and Intel Hybrid CPUs.
As a refresher, starting with the Intel 12th Generation (Alder Lake) CPU, a new hybrid big.LITTLE CPU architecture was introduced for consumer Intel CPUs. This new hybrid Intel CPU architecture integrates two types of CPU cores: Performance-cores (P-cores) and Efficiency-cores (E-cores) into the same physical CPU die. For more information about this new hybrid Intel CPU design, check out this resource HERE. The ESXi scheduler does not and has no current plans to support this new Intel Hybrid CPU architecture, especially as this type of architecture is nowhere to be found in traditional Enterprise datacenters and is only limited to Intel Consumer CPUs.
The current recommendation to work around the non-uniformity of the CPU cores is to either disable the E or P-cores within the system BIOS, thus making the system "uniform" and allowing ESXi to run like a normal x86 system. While you can apply a workaround to have ESXi ignore the non-uniformity of the CPU cores, in addition to the non-deterministic behaviors, random PSOD can also occur due to scheduling across two different types of cores.
I was curious to see whether applying ESXi CPU affinity on a VM using Intel Hybrid CPU Cores might yield a different outcome?
I first wanted to see if I could identify which CPU cores were P-cores versus E-cores with ESXi. For my experiment, I used the same Intel NUC 13 Pro which I had used for the VMware Workstation debugging, which has an an Intel i7 1360P (4 x P-Cores and 8 x E-Cores).
The observed behavior with VMware Workstation was that all P-cores (including hyperthreading) came came first, then followed by E-cores.
- Cores [ 0, 1, 2, 3, 4, 5, 6, 7] are all P-Cores (includes HT cores)
- Cores [8, 9, 10, 12, 13, 14, 15, 16] are all E-Cores
Unlike VMware Workstation, when ESXi observes non-uniform CPU cores, HT is automatically disabled by ESXi and thus we do not receive 2 x the P-Cores. To confirm whether ESXi has the same P and E-Core ordering behavior as VMware Workstation, I performed a simple test by iterating through each core and assigning it to a Windows VM to benchmark the performance using the popular CPU-Z utility. From this basic test, I was able to conclude that P-Cores were indeed ordered first followed by the E-Cores.
- Cores [ 0, 1, 2, 3] are all P-Cores (no HT cores)
- Cores [4, 5, 6, 7, 8, 9, 10, 11] are all E-Cores
Using this information, we can now create VMs and that are affinitized to either P-Cores or E-Cores to ensure consistent performance and hopefully avoid any inconsistent behaviors when schedule across different types of cores. If you are using a standalone non-managed ESXi host (e.g. no vCenter Server), you configure CPU affinity for a VM by using the ESXi Host Client and expanding the CPU configuration section.
Here is a VM configured with 2 x P-Cores which I have affinitized to Core 0 & 1
Here is a VM configured with 2 x E-Cores which I have affinitized to Core 4 & 5
If you have vCenter Server, it looks like CPU affinity was removed from the vSphere UI at some point and the only way to apply CPU affinity is by using the vSphere API. Below is a quick PowerCLI snippet for applying CPU affinity for a specific VM and this might even be better as it will allow you to easily apply the required affinity versus using the UI.
$vm = Get-VM "Win10-PCore-0-1"
$affinitySpec = New-Object VMware.Vim.VirtualMachineAffinityInfo
$affinitySpec.AffinitySet = @(0,1)
$spec = New-Object VMware.Vim.VirtualMachineConfigSpec
$spec.cpuAffinity = $affinitySpec
$task = $vm.ExtensionData.ReconfigVM_Task($spec)
$task1 = Get-Task -Id ("Task-$($task.value)")
$task1 | Wait-Task
As expected, we can see that the VM configured with 2 x P-Cores outperforms the VM configured with 2 x E-Cores.
For those looking to squeeze the most out of their hardware investments when using the new Intel Hybrid CPU Cores, there is at least an option to get consistent performance at the cost of manual CPU core assignment which could yield some CPU inefficiencies depending on how demanding your workloads are. I am curious to hear from the community on whether this is actually a feasible option for real world workloads since this was a pretty basic experiment and YMMV.
Hi William,
Thank you for valuable article. I'll definitelly test this out. I've aleady experimented with custom per-vm evc for nun-uniform esxi setup (based on your articles), and this looks like another level to squeeze performance out of homelabs.
Btw vsphere documentation for 8.0 mentions the affinity shall be set through vsphere client, so definitelly theres something whong somewhere.
I went with the 6c/12t i5-12400 when I rebuilt the lab this summer. Sidesteps the entire issue. Unfortunately, with Raptor Lake the best one can do is 4c/8t when avoiding big.LITTLE.
Not sure if it is still true (or if it was ever universally true across all OEMs) but I've seen reports that disabling P-Cores in BIOS still leaves 1 core enabled--you can disable all E-Cores but not all P-Cores.
I'm not sure that the power savings of disabling P-Cores would be all that remarkable in a homelab setting (in which the machine sits idle 99% of the time and c-states will ensure everything sleeps as much as possible), hence really the only incentive to disable cores is to avoid having to pin VMs to CPUs.
It's a shame VMware doesn't make minimal edits to ESXi to let somebody assign a CPUID mask to a VM and schedule it on a core that supports that mask. That feature already exists in the context of vMotion, so the only thing that's needed is to copy the same logic into the scheduler. It's certainly not a robust implementation and nowhere near good enough to offer formal support for BIG.little support in ESXi but it would be good enough to avoid having VMs crash.
How does ESXi on ARM handle this when you run on a raspberry pi?
"I'm not sure that the power savings of disabling P-Cores would be all that remarkable in a homelab setting"
I'm skeptical about the P-core/E-core thing period. Word on the street is the P-cores, when limited to select power states, will meet or exceed E-core performance-per-watt. The E-core marketing angle has been performance-per-watt but I'm not sure the reality matches up.
William can you please give us an honest feedback of the future of VMware. I see lots of articles about the negative things due to broadcon
Will, great write up. I've been playing with this sort of setup for awhile now. This actually made my cluster a ton more stable. Previously I had affinity rules for vms to separate hosts, 4 node cluster, and it seemed to stop the random psod. The biggest issue was aria operations. W/e.host that had it, would psod. This tweak stopped the psod. Do you ever think we will see an actual CPU scheduler for this sort of thing? I'd love to get back to vanilla setup without the overrides.
Hi William,
starting from 4th Xeon Scalable processor family the E & P core are available on Enterprise market.
Take a look on "Advanced Technologies" section for details:
https://ark.intel.com/content/www/us/en/ark/products/231750/intel-xeon-platinum-8468h-processor-105m-cache-2-10-ghz.html
Are not recognize correctly these high end processors by ESXi scheduler?
Beniamino
Hybrid CPU do not exists in Datacenter (eg Xeon SKUs). In future, it’ll either be all P or E, but not both hence it’s not an issue
This is available since 3rd gen, and unfortunatelly even it looks similar it's different technology, see: https://www.intel.com/content/www/us/en/support/articles/000094490/processors/intel-xeon-processors.html
As already mentioned, the Hybrid P/E Cores that's in Intel's Consumer segment are not expected to come into Datacenter SKUs and while there are some P/E-ness, they do NOT behave like consumer CPU and from an ESXi scheduler point of view, they will work like normal non-hybrid CPU. Furthermore, I know AMD have also announced they'd also be offering simliar type of P/E-like features in their datacenter offering, but again, they'll have consistent features, so from ESXi scheduler POV, they'll all look uniform and hence wouldn't have the problems observed with the consumer Intel Hybrid CPU
This is something different, low prio and high prio cores are basically the same architecture/capabilities with just different frequencies.
P/E cores in fact do have different architecture/capabilities.
Hi William how do you define the high and low priority core in 4th and 5th gen Xeon cpu. They seems to be different performance value. Will esxi code able to see them as same or different.
As mentioned a number of times, there are NO issues with ANY of the current and future Xeon CPU from Intel. The issues described here are ONLY found in Intel Consumer CPUs
Hi William,
Thanks for your tests and your work generally.
To conclude with ESXi architecture with Intel hybrid ; you recommend to use only P-core with HT (8 logicals cores) or both P/E-core without HT (12 logicals cores) ?
Your example is clear but I'm asking how to get the most out of my processor with several machines working at the time.
This is answered at the end of the article 🙂
Hi William,
Is there any news regarding the support of efficiency cores in upcoming updates in vSphere? I did some research myself, but unable to find anything regarding this.
Wonder if you know more ...
kind regards,
Oscar
No