Allow configurable CPU placement for KVM helper threads

My Windows 10 VM on Unraid 7.3.4 felt very slow despite having 28 vCPUs assigned. I found that one logical CPU repeatedly reached 100% while many others were lightly loaded. That was CPU 7 on the Unraid host, which maps to vCPU 2 in the Windows VM.

The kvm-nx-lpage-re helper thread shared CPU 7 with the vCPU thread. Together they used nearly 100% of one logical CPU. After I moved the helper outside the VM’s pinned CPU set, I measured average about 150% combined across two logical CPUs during a recovery burst. Windows felt much faster and more responsive in everyday use.

Could Unraid add a per-VM setting to choose where KVM helper threads run? I’d like to select either the last logical CPUs in the VM’s pinned set or a separate set of host CPUs outside that allocation.

I currently rely on a script to find the helper thread and move it again after every VM restart. A setting in the Unraid UI would help less technical users make better use of their hardware without having to diagnose CPU contention or maintain a custom script.

and here are statistics berofore and after

Unraid 7.3.4 | Windows 10 VM | Intel Xeon Gold 6148 | 28 vCPUs

BEFORE: KVM worker and guest vCPU 2 shared host logical CPU 7

Affinity checks:

pid 2303839's current affinity list: 7

pid 2303645's current affinity list: 7

Worker cgroup: vcpu2

Effective cpuset: 7

Selected top samples (COMMAND and %CPU transcribed from the original output):

Time kvm-nx-lpage-re CPU 2/KVM Combined

15:36:15 55.3% 41.9% 97.2%

15:36:17 56.8% 40.9% 97.7%

15:36:19 56.3% 41.3% 97.6%

15:36:21 55.4% 41.9% 97.3%

100% represents one host logical CPU. Both threads competed for the same CPU.

Unraid 7.3.4 | Windows 10 VM | Intel Xeon Gold 6148 | 28 vCPUs

AFTER: only the KVM worker was moved to the VM's emulator cgroup

Affinity change:

pid 2303839's current affinity list: 7

pid 2303839's new affinity list: 4,5,24,25

Guest vCPU 2 remained pinned to host logical CPU 7.

Selected top samples (thread CPU and CPU 7 idle transcribed from the original output):

Time kvm-nx-lpage-re CPU 2/KVM Combined Host CPU 7 idle

15:44:39 79.3% 60.3% 139.6% not sampled

15:48:51 98.8% 59.9% 158.7% 40.6%

15:48:53 99.2% 64.3% 163.5% 35.5%

15:48:59 99.2% 68.3% 167.5% 31.3%

15:49:03 84.7% 70.2% 154.9% 29.8%

Combined values are sums of the two thread readings. Values over 100%

show concurrent execution on separate logical CPUs. CPU 7 idle is

reported independently by top. Windows felt noticeably more responsive

in everyday use after the change. No frame-time benchmark was recorded.