Hi everyone,
I’m seeing a strange CPU scheduling/load balancing behavior with Proxmox VE 9.1 on Windows Server 2022 Terminal Server VMs.
Environment:
Proxmox VE 9.1
AMD EPYC 9375F hosts
Windows Server 2022 RDS/Terminal Server VMs
Multiple VMs affected
CPU type configured as either:
x86-64-v2-AES
EPYC-Genoa
Problem:
Inside the Windows guests, the second vCPU/core consistently shows noticeably higher load/utilization than all other cores. This happens across multiple VMs and is reproducible.
Interestingly, the issue disappears when I switch the VM CPU type to "host"
With host, CPU utilization is balanced normally across all vCPUs.
However, I would prefer to use an emulated CPU type (x86-64-v2-AES or EPYC-Genoa) because of migration compatibility and cluster flexibility.
What I already checked:
- VirtIO drivers up to date
- Different Multiqueue-values
- NUMA on/off tests
- Different vCPU counts
- Change of vCPU-Profile (NUMAstagic -> Closest)
- With/without VirtIO-RNG device
What confuses me is that I only found forum posts describing the opposite behavior — where host performs worse than generic/emulated CPU types.
Has anyone seen something similar?
Any ideas what could cause the uneven load distribution on the second vCPU specifically?
Thanks!
I’m seeing a strange CPU scheduling/load balancing behavior with Proxmox VE 9.1 on Windows Server 2022 Terminal Server VMs.
Environment:
Proxmox VE 9.1
AMD EPYC 9375F hosts
Windows Server 2022 RDS/Terminal Server VMs
Multiple VMs affected
CPU type configured as either:
x86-64-v2-AES
EPYC-Genoa
Problem:
Inside the Windows guests, the second vCPU/core consistently shows noticeably higher load/utilization than all other cores. This happens across multiple VMs and is reproducible.
Interestingly, the issue disappears when I switch the VM CPU type to "host"
With host, CPU utilization is balanced normally across all vCPUs.
However, I would prefer to use an emulated CPU type (x86-64-v2-AES or EPYC-Genoa) because of migration compatibility and cluster flexibility.
What I already checked:
- VirtIO drivers up to date
- Different Multiqueue-values
- NUMA on/off tests
- Different vCPU counts
- Change of vCPU-Profile (NUMAstagic -> Closest)
- With/without VirtIO-RNG device
What confuses me is that I only found forum posts describing the opposite behavior — where host performs worse than generic/emulated CPU types.
Has anyone seen something similar?
Any ideas what could cause the uneven load distribution on the second vCPU specifically?
Thanks!