This folder holds the Topology-aware scheduler plugin implementations based on Topology aware scheduler plugin based on NodeResourceTopology CRD. This plugin enables scheduling decisions based on worker node hardware topology overcoming the issue described here.
Document capturing the NodeResourceTopology API Custom Resource Definition Standard can be found here.
- 💡 Sample (for demonstrating and inspiring purpose)
- 👶 Alpha (used in companies for pilot projects)
- 👦 Beta (used in companies and developed actively)
- 👨 Stable (used in companies for production workloads)
In case the cumulative count of node resource allocatable appear to be the same for both the nodes in the cluster, topology aware scheduler plugin uses the CRD instance corresponding to the nodes to obtain the resource topology information to make a topology-aware scheduling decision.
Enable the "NodeResourceTopologyMatch" Fitler plugin via SchedulerConfigConfiguration.
apiVersion: kubescheduler.config.k8s.io/v1beta1
kind: KubeSchedulerConfiguration
leaderElection:
leaderElect: false
clientConnection:
kubeconfig: "/etc/kubernetes/scheduler.conf"
profiles:
- schedulerName: topo-aware-scheduler
plugins:
filter:
enabled:
- name: NodeResourceTopologyMatch
# optional plugin configs
pluginConfig:
- name: NodeResourceTopologyMatch
args:
kubeconfigpath: "/etc/kubernetes/scheduler.conf"
namespaces:
- default
- production
- test-namespaceLet us assume we have two nodes in a cluster deployed with sample-device-plugin with the hardware topology described by the diagram below:
The hardware topology corresponding to both the nodes is represented by the below CRD instances. These CRD instances are supposed to be created by Node Agents like Resource Topology Exporter (RTE) or Node feature Discovery (NFD). Please refer to issue Exposing Hardware Topology through CRDs in NFD and Design document which captures details of enhancing NFD to expose node resource topology through CRDs. Noderesourcetopology plugin works with namespaces, in this case each CRD could be namespace specific. The default namespace is used if it was omitted in the plugin's configuration.
# Worker Node A CRD spec
apiVersion: topology.node.k8s.io/v1alpha1
kind: NodeResourceTopology
metadata:
name: worker-node-A
namespace: test-namespace
topologyPolicies: ["SingleNUMANodeContainerLevel"]
zones:
- name: numa-node-0
type: Node
resources:
- name: cpu
capacity: 4
allocatable: 3
- name: example.com/deviceA
capacity: 1
allocatable: 1
- name: example.com/deviceB
capacity: 2
allocatable: 2
- name: numa-node-1
type: Node
resources:
- name: cpu
capacity: 4
allocatable: 3
- name: example.com/deviceA
capacity: 2
allocatable: 2
- name: example.com/deviceB
capacity: 1
allocatable: 1# Worker Node B CRD spec
apiVersion: topology.node.k8s.io/v1alpha1
kind: NodeResourceTopology
metadata:
name: worker-node-B
namespace: test-namespace
topologyPolicies: ["SingleNUMANodeContainerLevel"]
zones:
- name: numa-node-0
type: Node
resources:
- name: cpu
capacity: 4
allocatable: 3
- name: example.com/deviceA
capacity: 3
allocatable: 3
- name: numa-node-1
type: Node
resources:
- name: cpu
capacity: 4
allocatable: 3
- name: example.com/deviceB
capacity: 3
allocatable: 3-
Verify if the CRD has been created by running
-
In case NFD/RTE is deployed in the cluster ensure that the CRD and CRD instances are created by running
$ kubectl get noderesourcetopologies.topology.node.k8s.io
-
Alternatively, in case you are just interested in simply testing the scheduler plugin, use the manifest in the manifest directory to deploy the CRD and CRs as follows:
-
Deploy the Custom Resource Definition manifest
$ kubectl create -f crd.yaml
-
Check if the noderesourcetopologies.topology.node.k8s.io CRD is created
$ kubectl get crd $ kubectl get noderesourcetopologies.topology.node.k8s.io
-
Deploy the CRs representative of the hardware topology of the worker-node-A and worker-node-B:
$ kubectl create -f ns.yaml $ kubectl create -f worker-node-A.yaml $ kubectl create -f worker-node-B.yaml
NOTE: In case you are testing this demo by creating CRs manually, ensure that the names of the nodes in the cluster match the CR names.
-
-
-
Copy cluster kubeconfig file to /etc/kubernetes/scheduler.conf
-
Build the image locally
$ make local-image
-
Push the built image to the image registry:
$ docker push <IMAGE_REGISTRY>/scheduler-plugins/kube-scheduler:latest
-
Deploy the topology-aware scheduler plugin config
$ kubectl create -f scheduler-configmap.yaml
-
Deploy the Scheduler plugin
$ kubectl create -f cluster-role.yaml serviceaccount/topo-aware-scheduler created clusterrole.rbac.authorization.k8s.io/noderesourcetoplogy-handler created clusterrolebinding.rbac.authorization.k8s.io/topo-aware-scheduler-as-kube-scheduler created clusterrolebinding.rbac.authorization.k8s.io/my-scheduler-as-volume-scheduler created rolebinding.rbac.authorization.k8s.io/topo-aware-scheduler-as-kube-scheduler created clusterrolebinding.rbac.authorization.k8s.io/noderesourcetoplogy created $ kubectl create -f deploy.yaml deployment.apps/topo-aware-scheduler created
-
Check if the scheduler plugin is deployed correctly by running the following
$ kubectl get pods -n kube-system -o wide NAME READY STATUS RESTARTS AGE IP NODE NOMINATED NODE READINES topo-aware-scheduler-764c475854-vpmcw 1/1 Running 0 2s 10.244.0.14 kind-control-plane <none> <none> -
Deploy the pod to be scheduled with topology-aware scheduler plugin by populating the schedulerName: topo-aware-scheduler
apiVersion: apps/v1
kind: Deployment
metadata:
name: test-deployment
spec:
selector:
matchLabels:
name: test
template:
metadata:
labels:
name: test
spec:
schedulerName: topo-aware-scheduler
containers:
- name: test-deployment-1-container-1
image: nginx:1.7.9
imagePullPolicy: IfNotPresent
resources:
limits:
cpu: 1
memory: 100Mi
example.com/deviceA: 1
example.com/deviceB: 1
requests:
cpu: 1
memory: 100Mi
example.com/deviceA: 1
example.com/deviceB: 1$ kubectl create -f test-deployment.yaml
deployment.apps/test-deployment created-
The test-deployment pod should be scheduled on the worker-node-A node
$ kubectl get pods -o wide NAME READY STATUS RESTARTS AGE IP NODE NOMINATED NODE READINESS GATES device-plugin-a-ds-9bpsj 1/1 Running 0 3h13m 172.17.0.3 worker-node-B <none> <none> device-plugin-a-ds-dv55t 1/1 Running 0 3h13m 172.17.0.2 worker-node-A <none> <none> device-plugin-b-ds-8t7lh 1/1 Running 0 3h13m 172.17.0.2 worker-node-A <none> <none> device-plugin-b-ds-lt4pr 1/1 Running 0 3h13m 172.17.0.3 worker-node-B <none> <none> test-deployment-6dccf65ddb-pkg9j 1/1 Running 0 18s 172.17.0.2 worker-node-A <none> <none>