Horizontal Pod Autoscaling 可以根据 CPU 利用率自动伸缩一个 ReplicaSet、Deployment 或者中的 Pod 数量
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
| cat hpa-deploy.yaml
apiVersion: apps/v1 kind: Deployment metadata: name: hpa-web spec: selector: matchLabels: app: hpa-web replicas: 1 template: metadata: labels: app: hpa-web spec: containers: - name: hpa-web image: gcr.io/google_containers/hpa-example imagePullPolicy: IfNotPresent ports: - containerPort: 80 resources: limits: cpu: 200m memory: 200Mi requests: cpu: 100m memory: 100Mi --- apiVersion: v1 kind: Service metadata: name: hpa-web spec: type: ClusterIP selector: app: hpa-web ports: - name: http port: 80 targetPort: 80
|


创建 HPA 控制器 - 相关算法的详情请参阅这篇文档:http://git.k8s.io/community/contributors/design-proposals/horizontal-pod-autoscaler.md#autoscaling-algorithm
1
| kubectl autoscale deployment hpa-web --cpu-percent=50 --min=1 --max=10
|

增加负载,查看负载节点数目
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
| 两种方式:
1.命令行创建:
创建测试POD: kubectl run -i --tty test --image=busybox --image-pull-policy='IfNotPresent' /bin/sh
执行命令: while true; do wget -q -O- http://hpa-web.default.svc.cluster.local; done
2.副本控制器创建:
cat test.yaml
apiVersion: apps/v1 kind: Deployment metadata: name: ceshi spec: selector: matchLabels: app: ceshi replicas: 1 template: metadata: labels: app: ceshi spec: containers: - name: ceshi image: docker.io/busybox imagePullPolicy: IfNotPresent command: ['sh', '-c', 'while true; do wget -q -O- http://hpa-web.default.svc.cluster.local; done']
|



资源限制 - Pod
Kubernetes 对资源的限制实际上是通过 cgroup 来控制的,cgroup 是容器的一组用来控制内核如何运行进程的相关属性集合。针对内存、CPU 和各种设备都有对应的 cgroup
默认情况下,Pod 运行没有 CPU 和内存的限额。 这意味着系统中的任何 Pod 将能够像执行该 Pod 所在的节点一样,消耗足够多的 CPU 和内存 。一般会针对某些应用的 pod 资源进行资源限制,这个资源限制是通过resources 的 requests 和 limits 来实现
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
| spec: containers: - image: xxxx imagePullPolicy: Always name: auth ports: - containerPort: 8080 protocol: TCP resources: limits: cpu: "4" memory: 2Gi requests: cpu: 250m memory: 250Mi
|
requests 要分配的资源,limits 为最高请求的资源值。可以简单理解为初始值和最大值
资源限制 - 名称空间
1、计算资源配额
1 2 3 4 5 6 7 8 9 10 11 12
| apiVersion: v1 kind: ResourceQuota metadata: name: compute-resources namespace: test spec: hard: pods: "20" requests.cpu: "20" requests.memory: 100Gi limits.cpu: "40" limits.memory: 200Gi
|
2、配置对象数量配额限制
1 2 3 4 5 6 7 8 9 10 11 12
| apiVersion: v1 kind: ResourceQuota metadata: name: object-counts namespace: test spec: hard: configmaps: "10" persistentvolumeclaims: "4" ReplicaSet: "20" secrets: "10" services: "10"
|
3、配置 CPU 和 内存 LimitRange
1 2 3 4 5 6 7 8 9 10 11 12 13
| apiVersion: v1 kind: LimitRange metadata: name: mem-limit-range spec: limits: - default: memory: 50Gi cpu: 5 defaultRequest: memory: 1Gi cpu: 1 type: Container
|
default 即 limit 的值
defaultRequest 即 request 的值