search

torgeirl / boreas-scheduler

Custom Kubernetes scheduler that uses zephyrus2
MIT License
5 stars 3 forks source link

Node usage incorrect when Kubelet is configured with allocatable values #19

Closed torgeirl closed 1 year ago

torgeirl commented 1 year ago

There seems to be an issue with how nodes_usage() gather the resource usage on each worker node when Kubelet is configured to reserve CPU and RAM for system and Kube services even if BOREAS_SCHEDULER_RESERVED_KUBLET_CPU and BOREAS_SCHEDULER_RESERVED_KUBLET_RAM are set to 0.

For instance, a node with the following Kubelet configuration will be listed with 1800m CPU and 3589 MB RAM allocatable by kubectl describe node when its capacity is 2 CPUs and 4094 MB RAM:

kind: KubeletConfiguration
apiVersion: kubelet.config.k8s.io/v1beta1
systemReserved:
  cpu: 100m
  memory: 350M
kubeReserved:
  cpu: 100m
  memory: 50M
enforceNodeAllocatable:
- pods

Boreas currently lists its allocatable values at 1600m and 3487 RAM. Only 100m and 50Mi (52 MB) are requested by its kube-flannel pod, so there seems to be an additional 100m and 155 MB «missing».

torgeirl commented 1 year ago

Boreas now show 1700m CPU and 3537 MB RAM when only the kube-flannel pods are running.