noderesourcetopology

README

Overview

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.

Maturity Level

• 💡 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)

Tutorial

Expectation

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.

NOTE:

• NodeResourceTopology version v0.0.12 onwards, CRD has been changed from namespace to cluster scoped. Scheduler plugin version > v0.21.6 depends on NodeResourceTopology CRD v0.0.12 and the namespace field has been deprecated from the NodeResourceTopology scheduler config args.

Dependency:

• Scheduler plugin version <= v0.21.6 depends on the NodeResourceTopology CRD version v0.0.10.
• Scheduler plugin version > v0.21.6 depends on the NodeResourceTopology CRD version v0.0.12.

Config

Scheduler

Enable the "NodeResourceTopologyMatch" Filter and Score plugins via SchedulerConfigConfiguration.

apiVersion: kubescheduler.config.k8s.io/v1beta2
kind: KubeSchedulerConfiguration
leaderElection:
  leaderElect: false
clientConnection:
  kubeconfig: "/etc/kubernetes/scheduler.conf"
profiles:
- schedulerName: topo-aware-scheduler
  plugins:
    filter:
      enabled:
      - name: NodeResourceTopologyMatch
    score:
      enabled:
      - name: NodeResourceTopologyMatch
# optional plugin configs
  pluginConfig:
  - name: NodeResourceTopologyMatch
    args:
      # other strategies are MostAllocated and BalancedAllocation
      scoringStrategy:
        type: "LeastAllocated"

Cluster

The Topology-aware scheduler performs its decision over a number of node-specific hardware details or configuration settings which have node granularity (not at cluster granularity). Consistent settings across a set of nodes or all over the cluster is a fundamental prerequisite for the scheduler to work correctly. In other words, it is a prerequisite that a set of nodes share the same NUMA topology and kubelet configuration, at least for settings like topology and resource (cpu, memory, device) managers.

However, the scheduler has no means to enforce or even validate this prerequisite; for example the NodeResourceTopology CRD does not expose all the relevant fields, nor it should (it would be out of scope).

Hence, proper cluster configuration is expected from the cluster admins, or to other software components, like controllers or operators, outside of the scope here.

Should the cluster need to have different settings (e.g. topology manager) or NUMA topologies, we recommend to use the standard kubernetes tools to identify each set of nodes using affinity or also taints.

Demo

Let 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.

For configuring your cluster with NFD-topology updater, a software component in Node Feature Discovery which creates NodeResourceTopology CRs corresponding to nodes in the cluster follow the Quick Start guide here.

# Worker Node A CRD spec
apiVersion: topology.node.k8s.io/v1alpha1
kind: NodeResourceTopology
metadata:
  name: worker-node-A
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
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

  1. 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
     

  2. 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:

     1. Deploy the Custom Resource Definition manifest

        $ kubectl create -f crd.yaml
        

     2. Check if the noderesourcetopologies.topology.node.k8s.io CRD is created

         $ kubectl get crd
         $ kubectl get noderesourcetopologies.topology.node.k8s.io
        

     3. Deploy the CRs representative of the hardware topology of the worker-node-A and worker-node-B if CRs haven't been created using RTE or NFD as mentioned above:

         $ 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: quay.io/fromani/numalign
        imagePullPolicy: IfNotPresent
        command: ["/bin/sh", "-c"]
        args: [ "while true; do numalign; sleep 100000; done;" ]
        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>
  

Documentation

Index

• Constants
• func MakeTopologyResInfo(name, capacity, available string) topologyv1alpha1.ResourceInfo
• func New(args runtime.Object, handle framework.Handle) (framework.Plugin, error)
• type NUMANode
• type NUMANodeList
• type PolicyHandler
• type PolicyHandlerMap
• type TopologyMatch
  • func (tm *TopologyMatch) EventsToRegister() []framework.ClusterEvent
  • func (tm *TopologyMatch) Filter(ctx context.Context, cycleState *framework.CycleState, pod *v1.Pod, ...) *framework.Status
  • func (tm *TopologyMatch) Name() string
  • func (tm *TopologyMatch) Reserve(ctx context.Context, state *framework.CycleState, pod *corev1.Pod, ...) *framework.Status
  • func (tm *TopologyMatch) Score(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeName string) (int64, *framework.Status)
  • func (tm *TopologyMatch) ScoreExtensions() framework.ScoreExtensions
  • func (tm *TopologyMatch) Unreserve(ctx context.Context, state *framework.CycleState, pod *corev1.Pod, ...)

Constants

const (
	// Name is the name of the plugin used in the plugin registry and configurations.
	Name = "NodeResourceTopologyMatch"
)

Functions

func MakeTopologyResInfo

func MakeTopologyResInfo(name, capacity, available string) topologyv1alpha1.ResourceInfo

func New

func New(args runtime.Object, handle framework.Handle) (framework.Plugin, error)

New initializes a new plugin and returns it.

Types

type NUMANode

type NUMANode struct {
	NUMAID    int
	Resources v1.ResourceList
}

type NUMANodeList

type NUMANodeList []NUMANode

type PolicyHandler

type PolicyHandler func(pod *v1.Pod, zoneMap topologyv1alpha1.ZoneList) *framework.Status

type PolicyHandlerMap

type PolicyHandlerMap map[topologyv1alpha1.TopologyManagerPolicy]tmScopeHandler

type TopologyMatch

type TopologyMatch struct {
	// contains filtered or unexported fields
}

TopologyMatch plugin which run simplified version of TopologyManager's admit handler

func (*TopologyMatch) EventsToRegister

func (tm *TopologyMatch) EventsToRegister() []framework.ClusterEvent

EventsToRegister returns the possible events that may make a Pod failed by this plugin schedulable. NOTE: if in-place-update (KEP 1287) gets implemented, then PodUpdate event should be registered for this plugin since a Pod update may free up resources that make other Pods schedulable.

func (*TopologyMatch) Filter

func (tm *TopologyMatch) Filter(ctx context.Context, cycleState *framework.CycleState, pod *v1.Pod, nodeInfo *framework.NodeInfo) *framework.Status

Filter Now only single-numa-node supported

func (*TopologyMatch) Name

func (tm *TopologyMatch) Name() string

Name returns name of the plugin. It is used in logs, etc.

func (*TopologyMatch) Reserve

func (tm *TopologyMatch) Reserve(ctx context.Context, state *framework.CycleState, pod *corev1.Pod, nodeName string) *framework.Status

func (*TopologyMatch) Score

func (tm *TopologyMatch) Score(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeName string) (int64, *framework.Status)

func (*TopologyMatch) ScoreExtensions

func (tm *TopologyMatch) ScoreExtensions() framework.ScoreExtensions

func (*TopologyMatch) Unreserve

func (tm *TopologyMatch) Unreserve(ctx context.Context, state *framework.CycleState, pod *corev1.Pod, nodeName string)