Kubernetes: specify CPUs for cpumanager - kubernetes

Question

Is it possible to specify CPU ID list to the Kubernetes cpumanager? The goal is to make sure pods get CPUs from a single socket (0). I brought all the CPUs on the peer socket offline as mentioned here, for example:
$ echo 0 > /sys/devices/system/cpu/cpu5/online
After doing this, the Kubernetes master indeed sees the remaining online CPUs
kubectl describe node foo
Capacity:
cpu: 56 <<< socket 0 CPU count
ephemeral-storage: 958774760Ki
hugepages-1Gi: 120Gi
memory: 197524872Ki
pods: 110
Allocatable:
cpu: 54 <<< 2 system reserved CPUs
ephemeral-storage: 958774760Ki
hugepages-1Gi: 120Gi
memory: 71490952Ki
pods: 110
System Info:
Machine ID: 1155420082478559980231ba5bc0f6f2
System UUID: 4C4C4544-0044-4210-8031-C8C04F584B32
Boot ID: 7fa18227-748f-496c-968c-9fc82e21ecd5
Kernel Version: 4.4.13
OS Image: Ubuntu 16.04.4 LTS
Operating System: linux
Architecture: amd64
Container Runtime Version: docker://17.3.3
Kubelet Version: v1.11.1
Kube-Proxy Version: v1.11.1
However, cpumanager still seems to think there are 112 CPUs (socket0 + socket1).
cat /var/lib/kubelet/cpu_manager_state
{"policyName":"static","defaultCpuSet":"0-111"}
As a result, the kubelet system pods are throwing the following error:
kube-system kube-proxy-nk7gc 0/1 rpc error: code = Unknown desc = failed to update container "eb455f81a61b877eccda0d35eea7834e30f59615346140180f08077f64896760": Error response from daemon: Requested CPUs are not available - requested 0-111, available: 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110 762 36d <IP address> foo <none>

Answer 1

I was able to get this working. Posting this as an answer so that someone in need might benefit.
It appears the CPU set is read from /var/lib/kubelet/cpu_manager_state file and it is not updated across kubelet restarts. So this file needs to be removed before restarting kubelet.
The following worked for me:
# On a running worker node, bring desired CPUs offline. (run as root)
$ cpu_list=`lscpu | grep "NUMA node1 CPU(s)" | awk '{print $4}'`
$ chcpu -d $cpu_list
$ rm -f /var/lib/kubelet/cpu_manager_state
$ systemctl restart kubelet.service
# Check the CPU set seen by the CPU manager
$ cat /var/lib/kubelet/cpu_manager_state
# Try creating pods and check the syslog:
Dec 3 14:36:05 k8-2-w1 kubelet[8070]: I1203 14:36:05.122466 8070 state_mem.go:84] [cpumanager] updated default cpuset: "0,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110"
Dec 3 14:36:05 k8-2-w1 kubelet[8070]: I1203 14:36:05.122643 8070 policy_static.go:198] [cpumanager] allocateCPUs: returning "2,4,6,8,58,60,62,64"
Dec 3 14:36:05 k8-2-w1 kubelet[8070]: I1203 14:36:05.122660 8070 state_mem.go:76] [cpumanager] updated desired cpuset (container id: 356939cdf32d0f719e83b0029a018a2ca2c349fc0bdc1004da5d842e357c503a, cpuset: "2,4,6,8,58,60,62,64")
I have reported a bug here as I think the CPU set should be updated after kubelet restarts.