mirror of
https://github.com/optim-enterprises-bv/kubernetes.git
synced 2025-10-31 18:28:13 +00:00
Remove remants of broken stuff - nvidia/autoscaling
Signed-off-by: Davanum Srinivas <davanum@gmail.com>
This commit is contained in:
Davanum Srinivas
@@ -1,57 +0,0 @@
|
||||
apiVersion: apps/v1
|
||||
kind: DaemonSet
|
||||
metadata:
|
||||
name: nvidia-gpu-device-plugin
|
||||
namespace: kube-system
|
||||
labels:
|
||||
k8s-app: nvidia-gpu-device-plugin
|
||||
addonmanager.kubernetes.io/mode: EnsureExists
|
||||
spec:
|
||||
selector:
|
||||
matchLabels:
|
||||
k8s-app: nvidia-gpu-device-plugin
|
||||
template:
|
||||
metadata:
|
||||
labels:
|
||||
k8s-app: nvidia-gpu-device-plugin
|
||||
spec:
|
||||
priorityClassName: system-node-critical
|
||||
affinity:
|
||||
nodeAffinity:
|
||||
requiredDuringSchedulingIgnoredDuringExecution:
|
||||
nodeSelectorTerms:
|
||||
- matchExpressions:
|
||||
- key: cloud.google.com/gke-accelerator
|
||||
operator: Exists
|
||||
tolerations:
|
||||
- operator: "Exists"
|
||||
effect: "NoExecute"
|
||||
- operator: "Exists"
|
||||
effect: "NoSchedule"
|
||||
volumes:
|
||||
- name: device-plugin
|
||||
hostPath:
|
||||
path: /var/lib/kubelet/device-plugins
|
||||
- name: dev
|
||||
hostPath:
|
||||
path: /dev
|
||||
containers:
|
||||
- image: "registry.k8s.io/nvidia-gpu-device-plugin@sha256:4b036e8844920336fa48f36edeb7d4398f426d6a934ba022848deed2edbf09aa"
|
||||
command: ["/usr/bin/nvidia-gpu-device-plugin", "-logtostderr"]
|
||||
name: nvidia-gpu-device-plugin
|
||||
resources:
|
||||
requests:
|
||||
cpu: 50m
|
||||
memory: 10Mi
|
||||
limits:
|
||||
cpu: 50m
|
||||
memory: 10Mi
|
||||
securityContext:
|
||||
privileged: true
|
||||
volumeMounts:
|
||||
- name: device-plugin
|
||||
mountPath: /device-plugin
|
||||
- name: dev
|
||||
mountPath: /dev
|
||||
updateStrategy:
|
||||
type: RollingUpdate
|
||||
@@ -2943,9 +2943,6 @@ EOF
|
||||
sed -i -e "s@{{ metrics_server_memory_per_node }}@${metrics_server_memory_per_node}@g" "${metrics_server_yaml}"
|
||||
sed -i -e "s@{{ metrics_server_min_cluster_size }}@${metrics_server_min_cluster_size}@g" "${metrics_server_yaml}"
|
||||
fi
|
||||
if [[ "${ENABLE_NVIDIA_GPU_DEVICE_PLUGIN:-}" == "true" ]]; then
|
||||
setup-addon-manifests "addons" "device-plugins/nvidia-gpu"
|
||||
fi
|
||||
# Setting up the konnectivity-agent daemonset
|
||||
if [[ "${RUN_KONNECTIVITY_PODS:-false}" == "true" ]]; then
|
||||
setup-addon-manifests "addons" "konnectivity-agent"
|
||||
|
||||
@@ -1512,11 +1512,6 @@ EOF
|
||||
if [ -n "${CLUSTER_SIGNING_DURATION:-}" ]; then
|
||||
cat >>"$file" <<EOF
|
||||
CLUSTER_SIGNING_DURATION: $(yaml-quote "${CLUSTER_SIGNING_DURATION}")
|
||||
EOF
|
||||
fi
|
||||
if [[ "${NODE_ACCELERATORS:-}" == *"type=nvidia"* ]]; then
|
||||
cat >>"$file" <<EOF
|
||||
ENABLE_NVIDIA_GPU_DEVICE_PLUGIN: $(yaml-quote "true")
|
||||
EOF
|
||||
fi
|
||||
if [ -n "${ADDON_MANAGER_LEADER_ELECTION:-}" ]; then
|
||||
|
||||
@@ -1,125 +0,0 @@
|
||||
/*
|
||||
Copyright 2017 The Kubernetes Authors.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
*/
|
||||
|
||||
package autoscaling
|
||||
|
||||
import (
|
||||
"context"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
v1 "k8s.io/api/core/v1"
|
||||
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
|
||||
"k8s.io/kubernetes/test/e2e/feature"
|
||||
"k8s.io/kubernetes/test/e2e/framework"
|
||||
e2eautoscaling "k8s.io/kubernetes/test/e2e/framework/autoscaling"
|
||||
e2enode "k8s.io/kubernetes/test/e2e/framework/node"
|
||||
e2eskipper "k8s.io/kubernetes/test/e2e/framework/skipper"
|
||||
admissionapi "k8s.io/pod-security-admission/api"
|
||||
|
||||
"github.com/onsi/ginkgo/v2"
|
||||
"github.com/onsi/gomega"
|
||||
"github.com/onsi/gomega/gmeasure"
|
||||
)
|
||||
|
||||
var _ = SIGDescribe(feature.ClusterSizeAutoscalingScaleUp, framework.WithSlow(), "Autoscaling", func() {
|
||||
f := framework.NewDefaultFramework("autoscaling")
|
||||
f.NamespacePodSecurityLevel = admissionapi.LevelPrivileged
|
||||
var experiment *gmeasure.Experiment
|
||||
|
||||
ginkgo.Describe("Autoscaling a service", func() {
|
||||
ginkgo.BeforeEach(func(ctx context.Context) {
|
||||
// Check if Cloud Autoscaler is enabled by trying to get its ConfigMap.
|
||||
_, err := f.ClientSet.CoreV1().ConfigMaps("kube-system").Get(ctx, "cluster-autoscaler-status", metav1.GetOptions{})
|
||||
if err != nil {
|
||||
e2eskipper.Skipf("test expects Cluster Autoscaler to be enabled")
|
||||
}
|
||||
experiment = gmeasure.NewExperiment("Autoscaling a service")
|
||||
ginkgo.AddReportEntry(experiment.Name, experiment)
|
||||
})
|
||||
|
||||
ginkgo.Context("from 1 pod and 3 nodes to 8 pods and >=4 nodes", func() {
|
||||
const nodesNum = 3 // Expect there to be 3 nodes before and after the test.
|
||||
var nodeGroupName string // Set by BeforeEach, used by AfterEach to scale this node group down after the test.
|
||||
var nodes *v1.NodeList // Set by BeforeEach, used by Measure to calculate CPU request based on node's sizes.
|
||||
|
||||
ginkgo.BeforeEach(func(ctx context.Context) {
|
||||
// Make sure there is only 1 node group, otherwise this test becomes useless.
|
||||
nodeGroups := strings.Split(framework.TestContext.CloudConfig.NodeInstanceGroup, ",")
|
||||
if len(nodeGroups) != 1 {
|
||||
e2eskipper.Skipf("test expects 1 node group, found %d", len(nodeGroups))
|
||||
}
|
||||
nodeGroupName = nodeGroups[0]
|
||||
|
||||
// Make sure the node group has exactly 'nodesNum' nodes, otherwise this test becomes useless.
|
||||
nodeGroupSize, err := framework.GroupSize(nodeGroupName)
|
||||
framework.ExpectNoError(err)
|
||||
if nodeGroupSize != nodesNum {
|
||||
e2eskipper.Skipf("test expects %d nodes, found %d", nodesNum, nodeGroupSize)
|
||||
}
|
||||
|
||||
// Make sure all nodes are schedulable, otherwise we are in some kind of a problem state.
|
||||
nodes, err = e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
|
||||
framework.ExpectNoError(err)
|
||||
gomega.Expect(nodes.Items).To(gomega.HaveLen(nodeGroupSize), "not all nodes are schedulable")
|
||||
})
|
||||
|
||||
ginkgo.AfterEach(func(ctx context.Context) {
|
||||
// Attempt cleanup only if a node group was targeted for scale up.
|
||||
// Otherwise the test was probably skipped and we'll get a gcloud error due to invalid parameters.
|
||||
if len(nodeGroupName) > 0 {
|
||||
// Scale down back to only 'nodesNum' nodes, as expected at the start of the test.
|
||||
framework.ExpectNoError(framework.ResizeGroup(nodeGroupName, nodesNum))
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, f.ClientSet, nodesNum, 15*time.Minute))
|
||||
}
|
||||
})
|
||||
|
||||
ginkgo.It("takes less than 15 minutes", func(ctx context.Context) {
|
||||
// Measured over multiple samples, scaling takes 10 +/- 2 minutes, so 15 minutes should be fully sufficient.
|
||||
const timeToWait = 15 * time.Minute
|
||||
|
||||
// Calculate the CPU request of the service.
|
||||
// This test expects that 8 pods will not fit in 'nodesNum' nodes, but will fit in >='nodesNum'+1 nodes.
|
||||
// Make it so that 'nodesNum' pods fit perfectly per node.
|
||||
nodeCpus := nodes.Items[0].Status.Allocatable[v1.ResourceCPU]
|
||||
nodeCPUMillis := (&nodeCpus).MilliValue()
|
||||
cpuRequestMillis := int64(nodeCPUMillis / nodesNum)
|
||||
|
||||
// Start the service we want to scale and wait for it to be up and running.
|
||||
nodeMemoryBytes := nodes.Items[0].Status.Allocatable[v1.ResourceMemory]
|
||||
nodeMemoryMB := (&nodeMemoryBytes).Value() / 1024 / 1024
|
||||
memRequestMB := nodeMemoryMB / 10 // Ensure each pod takes not more than 10% of node's allocatable memory.
|
||||
replicas := 1
|
||||
resourceConsumer := e2eautoscaling.NewDynamicResourceConsumer(ctx, "resource-consumer", f.Namespace.Name, e2eautoscaling.KindDeployment, replicas, 0, 0, 0, cpuRequestMillis, memRequestMB, f.ClientSet, f.ScalesGetter, e2eautoscaling.Disable, e2eautoscaling.Idle)
|
||||
ginkgo.DeferCleanup(resourceConsumer.CleanUp)
|
||||
resourceConsumer.WaitForReplicas(ctx, replicas, 1*time.Minute) // Should finish ~immediately, so 1 minute is more than enough.
|
||||
|
||||
// Enable Horizontal Pod Autoscaler with 50% target utilization and
|
||||
// scale up the CPU usage to trigger autoscaling to 8 pods for target to be satisfied.
|
||||
targetCPUUtilizationPercent := int32(50)
|
||||
hpa := e2eautoscaling.CreateCPUResourceHorizontalPodAutoscaler(ctx, resourceConsumer, targetCPUUtilizationPercent, 1, 10)
|
||||
ginkgo.DeferCleanup(e2eautoscaling.DeleteHorizontalPodAutoscaler, resourceConsumer, hpa.Name)
|
||||
cpuLoad := 8 * cpuRequestMillis * int64(targetCPUUtilizationPercent) / 100 // 8 pods utilized to the target level
|
||||
resourceConsumer.ConsumeCPU(int(cpuLoad))
|
||||
|
||||
// Measure the time it takes for the service to scale to 8 pods with 50% CPU utilization each.
|
||||
experiment.SampleDuration("total scale-up time", func(idx int) {
|
||||
resourceConsumer.WaitForReplicas(ctx, 8, timeToWait)
|
||||
}, gmeasure.SamplingConfig{N: 1})
|
||||
}) // Increase to run the test more than once.
|
||||
})
|
||||
})
|
||||
})
|
||||
@@ -1,531 +0,0 @@
|
||||
/*
|
||||
Copyright 2016 The Kubernetes Authors.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
*/
|
||||
|
||||
package autoscaling
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"math"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
v1 "k8s.io/api/core/v1"
|
||||
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
|
||||
"k8s.io/apimachinery/pkg/fields"
|
||||
"k8s.io/apimachinery/pkg/types"
|
||||
"k8s.io/apimachinery/pkg/util/strategicpatch"
|
||||
clientset "k8s.io/client-go/kubernetes"
|
||||
"k8s.io/klog/v2"
|
||||
"k8s.io/kubernetes/test/e2e/feature"
|
||||
"k8s.io/kubernetes/test/e2e/framework"
|
||||
e2enode "k8s.io/kubernetes/test/e2e/framework/node"
|
||||
e2erc "k8s.io/kubernetes/test/e2e/framework/rc"
|
||||
e2eskipper "k8s.io/kubernetes/test/e2e/framework/skipper"
|
||||
testutils "k8s.io/kubernetes/test/utils"
|
||||
imageutils "k8s.io/kubernetes/test/utils/image"
|
||||
admissionapi "k8s.io/pod-security-admission/api"
|
||||
|
||||
"github.com/onsi/ginkgo/v2"
|
||||
"github.com/onsi/gomega"
|
||||
)
|
||||
|
||||
const (
|
||||
memoryReservationTimeout = 5 * time.Minute
|
||||
largeResizeTimeout = 8 * time.Minute
|
||||
largeScaleUpTimeout = 10 * time.Minute
|
||||
maxNodes = 1000
|
||||
)
|
||||
|
||||
type clusterPredicates struct {
|
||||
nodes int
|
||||
}
|
||||
|
||||
type scaleUpTestConfig struct {
|
||||
initialNodes int
|
||||
initialPods int
|
||||
extraPods *testutils.RCConfig
|
||||
expectedResult *clusterPredicates
|
||||
}
|
||||
|
||||
var _ = SIGDescribe("Cluster size autoscaler scalability", framework.WithSlow(), func() {
|
||||
f := framework.NewDefaultFramework("autoscaling")
|
||||
f.NamespacePodSecurityLevel = admissionapi.LevelPrivileged
|
||||
var c clientset.Interface
|
||||
var nodeCount int
|
||||
var coresPerNode int
|
||||
var memCapacityMb int
|
||||
var originalSizes map[string]int
|
||||
var sum int
|
||||
|
||||
ginkgo.BeforeEach(func(ctx context.Context) {
|
||||
e2eskipper.SkipUnlessProviderIs("gce", "gke", "kubemark")
|
||||
|
||||
// Check if Cloud Autoscaler is enabled by trying to get its ConfigMap.
|
||||
_, err := f.ClientSet.CoreV1().ConfigMaps("kube-system").Get(ctx, "cluster-autoscaler-status", metav1.GetOptions{})
|
||||
if err != nil {
|
||||
e2eskipper.Skipf("test expects Cluster Autoscaler to be enabled")
|
||||
}
|
||||
|
||||
c = f.ClientSet
|
||||
if originalSizes == nil {
|
||||
originalSizes = make(map[string]int)
|
||||
sum = 0
|
||||
for _, mig := range strings.Split(framework.TestContext.CloudConfig.NodeInstanceGroup, ",") {
|
||||
size, err := framework.GroupSize(mig)
|
||||
framework.ExpectNoError(err)
|
||||
ginkgo.By(fmt.Sprintf("Initial size of %s: %d", mig, size))
|
||||
originalSizes[mig] = size
|
||||
sum += size
|
||||
}
|
||||
}
|
||||
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, sum, scaleUpTimeout))
|
||||
|
||||
nodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
|
||||
framework.ExpectNoError(err)
|
||||
nodeCount = len(nodes.Items)
|
||||
cpu := nodes.Items[0].Status.Capacity[v1.ResourceCPU]
|
||||
mem := nodes.Items[0].Status.Capacity[v1.ResourceMemory]
|
||||
coresPerNode = int((&cpu).MilliValue() / 1000)
|
||||
memCapacityMb = int((&mem).Value() / 1024 / 1024)
|
||||
|
||||
gomega.Expect(nodeCount).To(gomega.Equal(sum))
|
||||
|
||||
if framework.ProviderIs("gke") {
|
||||
val, err := isAutoscalerEnabled(3)
|
||||
framework.ExpectNoError(err)
|
||||
if !val {
|
||||
err = enableAutoscaler("default-pool", 3, 5)
|
||||
framework.ExpectNoError(err)
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
ginkgo.AfterEach(func(ctx context.Context) {
|
||||
ginkgo.By(fmt.Sprintf("Restoring initial size of the cluster"))
|
||||
setMigSizes(originalSizes)
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount, scaleDownTimeout))
|
||||
nodes, err := c.CoreV1().Nodes().List(ctx, metav1.ListOptions{})
|
||||
framework.ExpectNoError(err)
|
||||
s := time.Now()
|
||||
makeSchedulableLoop:
|
||||
for start := time.Now(); time.Since(start) < makeSchedulableTimeout; time.Sleep(makeSchedulableDelay) {
|
||||
for _, n := range nodes.Items {
|
||||
err = makeNodeSchedulable(ctx, c, &n, true)
|
||||
switch err.(type) {
|
||||
case CriticalAddonsOnlyError:
|
||||
continue makeSchedulableLoop
|
||||
default:
|
||||
framework.ExpectNoError(err)
|
||||
}
|
||||
}
|
||||
break
|
||||
}
|
||||
klog.Infof("Made nodes schedulable again in %v", time.Since(s).String())
|
||||
})
|
||||
|
||||
f.It("should scale up at all", feature.ClusterAutoscalerScalability1, func(ctx context.Context) {
|
||||
perNodeReservation := int(float64(memCapacityMb) * 0.95)
|
||||
replicasPerNode := 10
|
||||
|
||||
additionalNodes := maxNodes - nodeCount
|
||||
replicas := additionalNodes * replicasPerNode
|
||||
additionalReservation := additionalNodes * perNodeReservation
|
||||
|
||||
// saturate cluster
|
||||
reservationCleanup := ReserveMemory(ctx, f, "some-pod", nodeCount*2, nodeCount*perNodeReservation, true, memoryReservationTimeout)
|
||||
defer reservationCleanup()
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
|
||||
|
||||
// configure pending pods & expected scale up
|
||||
rcConfig := reserveMemoryRCConfig(f, "extra-pod-1", replicas, additionalReservation, largeScaleUpTimeout)
|
||||
expectedResult := createClusterPredicates(nodeCount + additionalNodes)
|
||||
config := createScaleUpTestConfig(nodeCount, nodeCount, rcConfig, expectedResult)
|
||||
|
||||
// run test
|
||||
testCleanup := simpleScaleUpTest(ctx, f, config)
|
||||
defer testCleanup()
|
||||
})
|
||||
|
||||
f.It("should scale up twice", feature.ClusterAutoscalerScalability2, func(ctx context.Context) {
|
||||
perNodeReservation := int(float64(memCapacityMb) * 0.95)
|
||||
replicasPerNode := 10
|
||||
additionalNodes1 := int(math.Ceil(0.7 * maxNodes))
|
||||
additionalNodes2 := int(math.Ceil(0.25 * maxNodes))
|
||||
if additionalNodes1+additionalNodes2 > maxNodes {
|
||||
additionalNodes2 = maxNodes - additionalNodes1
|
||||
}
|
||||
|
||||
replicas1 := additionalNodes1 * replicasPerNode
|
||||
replicas2 := additionalNodes2 * replicasPerNode
|
||||
|
||||
klog.Infof("cores per node: %v", coresPerNode)
|
||||
|
||||
// saturate cluster
|
||||
initialReplicas := nodeCount
|
||||
reservationCleanup := ReserveMemory(ctx, f, "some-pod", initialReplicas, nodeCount*perNodeReservation, true, memoryReservationTimeout)
|
||||
defer reservationCleanup()
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
|
||||
|
||||
klog.Infof("Reserved successfully")
|
||||
|
||||
// configure pending pods & expected scale up #1
|
||||
rcConfig := reserveMemoryRCConfig(f, "extra-pod-1", replicas1, additionalNodes1*perNodeReservation, largeScaleUpTimeout)
|
||||
expectedResult := createClusterPredicates(nodeCount + additionalNodes1)
|
||||
config := createScaleUpTestConfig(nodeCount, nodeCount, rcConfig, expectedResult)
|
||||
|
||||
// run test #1
|
||||
tolerateUnreadyNodes := additionalNodes1 / 20
|
||||
tolerateUnreadyPods := (initialReplicas + replicas1) / 20
|
||||
testCleanup1 := simpleScaleUpTestWithTolerance(ctx, f, config, tolerateUnreadyNodes, tolerateUnreadyPods)
|
||||
defer testCleanup1()
|
||||
|
||||
klog.Infof("Scaled up once")
|
||||
|
||||
// configure pending pods & expected scale up #2
|
||||
rcConfig2 := reserveMemoryRCConfig(f, "extra-pod-2", replicas2, additionalNodes2*perNodeReservation, largeScaleUpTimeout)
|
||||
expectedResult2 := createClusterPredicates(nodeCount + additionalNodes1 + additionalNodes2)
|
||||
config2 := createScaleUpTestConfig(nodeCount+additionalNodes1, nodeCount+additionalNodes2, rcConfig2, expectedResult2)
|
||||
|
||||
// run test #2
|
||||
tolerateUnreadyNodes = maxNodes / 20
|
||||
tolerateUnreadyPods = (initialReplicas + replicas1 + replicas2) / 20
|
||||
testCleanup2 := simpleScaleUpTestWithTolerance(ctx, f, config2, tolerateUnreadyNodes, tolerateUnreadyPods)
|
||||
defer testCleanup2()
|
||||
|
||||
klog.Infof("Scaled up twice")
|
||||
})
|
||||
|
||||
f.It("should scale down empty nodes", feature.ClusterAutoscalerScalability3, func(ctx context.Context) {
|
||||
perNodeReservation := int(float64(memCapacityMb) * 0.7)
|
||||
replicas := int(math.Ceil(maxNodes * 0.7))
|
||||
totalNodes := maxNodes
|
||||
|
||||
// resize cluster to totalNodes
|
||||
newSizes := map[string]int{
|
||||
anyKey(originalSizes): totalNodes,
|
||||
}
|
||||
setMigSizes(newSizes)
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, f.ClientSet, totalNodes, largeResizeTimeout))
|
||||
|
||||
// run replicas
|
||||
rcConfig := reserveMemoryRCConfig(f, "some-pod", replicas, replicas*perNodeReservation, largeScaleUpTimeout)
|
||||
expectedResult := createClusterPredicates(totalNodes)
|
||||
config := createScaleUpTestConfig(totalNodes, totalNodes, rcConfig, expectedResult)
|
||||
tolerateUnreadyNodes := totalNodes / 10
|
||||
tolerateUnreadyPods := replicas / 10
|
||||
testCleanup := simpleScaleUpTestWithTolerance(ctx, f, config, tolerateUnreadyNodes, tolerateUnreadyPods)
|
||||
defer testCleanup()
|
||||
|
||||
// check if empty nodes are scaled down
|
||||
framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
|
||||
func(size int) bool {
|
||||
return size <= replicas+3 // leaving space for non-evictable kube-system pods
|
||||
}, scaleDownTimeout))
|
||||
})
|
||||
|
||||
f.It("should scale down underutilized nodes", feature.ClusterAutoscalerScalability4, func(ctx context.Context) {
|
||||
perPodReservation := int(float64(memCapacityMb) * 0.01)
|
||||
// underutilizedNodes are 10% full
|
||||
underutilizedPerNodeReplicas := 10
|
||||
// fullNodes are 70% full
|
||||
fullPerNodeReplicas := 70
|
||||
totalNodes := maxNodes
|
||||
underutilizedRatio := 0.3
|
||||
maxDelta := 30
|
||||
|
||||
// resize cluster to totalNodes
|
||||
newSizes := map[string]int{
|
||||
anyKey(originalSizes): totalNodes,
|
||||
}
|
||||
setMigSizes(newSizes)
|
||||
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, f.ClientSet, totalNodes, largeResizeTimeout))
|
||||
|
||||
// annotate all nodes with no-scale-down
|
||||
ScaleDownDisabledKey := "cluster-autoscaler.kubernetes.io/scale-down-disabled"
|
||||
|
||||
nodes, err := f.ClientSet.CoreV1().Nodes().List(ctx, metav1.ListOptions{
|
||||
FieldSelector: fields.Set{
|
||||
"spec.unschedulable": "false",
|
||||
}.AsSelector().String(),
|
||||
})
|
||||
|
||||
framework.ExpectNoError(err)
|
||||
framework.ExpectNoError(addAnnotation(ctx, f, nodes.Items, ScaleDownDisabledKey, "true"))
|
||||
|
||||
// distribute pods using replication controllers taking up space that should
|
||||
// be empty after pods are distributed
|
||||
underutilizedNodesNum := int(float64(maxNodes) * underutilizedRatio)
|
||||
fullNodesNum := totalNodes - underutilizedNodesNum
|
||||
|
||||
podDistribution := []podBatch{
|
||||
{numNodes: fullNodesNum, podsPerNode: fullPerNodeReplicas},
|
||||
{numNodes: underutilizedNodesNum, podsPerNode: underutilizedPerNodeReplicas}}
|
||||
|
||||
distributeLoad(ctx, f, f.Namespace.Name, "10-70", podDistribution, perPodReservation,
|
||||
int(0.95*float64(memCapacityMb)), map[string]string{}, largeScaleUpTimeout)
|
||||
|
||||
// enable scale down again
|
||||
framework.ExpectNoError(addAnnotation(ctx, f, nodes.Items, ScaleDownDisabledKey, "false"))
|
||||
|
||||
// wait for scale down to start. Node deletion takes a long time, so we just
|
||||
// wait for maximum of 30 nodes deleted
|
||||
nodesToScaleDownCount := int(float64(totalNodes) * 0.1)
|
||||
if nodesToScaleDownCount > maxDelta {
|
||||
nodesToScaleDownCount = maxDelta
|
||||
}
|
||||
expectedSize := totalNodes - nodesToScaleDownCount
|
||||
timeout := time.Duration(nodesToScaleDownCount)*time.Minute + scaleDownTimeout
|
||||
framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet, func(size int) bool {
|
||||
return size <= expectedSize
|
||||
}, timeout))
|
||||
})
|
||||
|
||||
f.It("shouldn't scale down with underutilized nodes due to host port conflicts", feature.ClusterAutoscalerScalability5, func(ctx context.Context) {
|
||||
fullReservation := int(float64(memCapacityMb) * 0.9)
|
||||
hostPortPodReservation := int(float64(memCapacityMb) * 0.3)
|
||||
totalNodes := maxNodes
|
||||
reservedPort := 4321
|
||||
|
||||
// resize cluster to totalNodes
|
||||
newSizes := map[string]int{
|
||||
anyKey(originalSizes): totalNodes,
|
||||
}
|
||||
setMigSizes(newSizes)
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, f.ClientSet, totalNodes, largeResizeTimeout))
|
||||
divider := int(float64(totalNodes) * 0.7)
|
||||
fullNodesCount := divider
|
||||
underutilizedNodesCount := totalNodes - fullNodesCount
|
||||
|
||||
ginkgo.By("Reserving full nodes")
|
||||
// run RC1 w/o host port
|
||||
cleanup := ReserveMemory(ctx, f, "filling-pod", fullNodesCount, fullNodesCount*fullReservation, true, largeScaleUpTimeout*2)
|
||||
defer cleanup()
|
||||
|
||||
ginkgo.By("Reserving host ports on remaining nodes")
|
||||
// run RC2 w/ host port
|
||||
ginkgo.DeferCleanup(createHostPortPodsWithMemory, f, "underutilizing-host-port-pod", underutilizedNodesCount, reservedPort, underutilizedNodesCount*hostPortPodReservation, largeScaleUpTimeout)
|
||||
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
|
||||
// wait and check scale down doesn't occur
|
||||
ginkgo.By(fmt.Sprintf("Sleeping %v minutes...", scaleDownTimeout.Minutes()))
|
||||
time.Sleep(scaleDownTimeout)
|
||||
|
||||
ginkgo.By("Checking if the number of nodes is as expected")
|
||||
nodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
|
||||
framework.ExpectNoError(err)
|
||||
klog.Infof("Nodes: %v, expected: %v", len(nodes.Items), totalNodes)
|
||||
gomega.Expect(nodes.Items).To(gomega.HaveLen(totalNodes))
|
||||
})
|
||||
|
||||
f.It("CA ignores unschedulable pods while scheduling schedulable pods", feature.ClusterAutoscalerScalability6, func(ctx context.Context) {
|
||||
// Start a number of pods saturating existing nodes.
|
||||
perNodeReservation := int(float64(memCapacityMb) * 0.80)
|
||||
replicasPerNode := 10
|
||||
initialPodReplicas := nodeCount * replicasPerNode
|
||||
initialPodsTotalMemory := nodeCount * perNodeReservation
|
||||
reservationCleanup := ReserveMemory(ctx, f, "initial-pod", initialPodReplicas, initialPodsTotalMemory, true /* wait for pods to run */, memoryReservationTimeout)
|
||||
ginkgo.DeferCleanup(reservationCleanup)
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
|
||||
|
||||
// Configure a number of unschedulable pods.
|
||||
unschedulableMemReservation := memCapacityMb * 2
|
||||
unschedulablePodReplicas := 1000
|
||||
totalMemReservation := unschedulableMemReservation * unschedulablePodReplicas
|
||||
timeToWait := 5 * time.Minute
|
||||
podsConfig := reserveMemoryRCConfig(f, "unschedulable-pod", unschedulablePodReplicas, totalMemReservation, timeToWait)
|
||||
_ = e2erc.RunRC(ctx, *podsConfig) // Ignore error (it will occur because pods are unschedulable)
|
||||
ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, podsConfig.Name)
|
||||
|
||||
// Ensure that no new nodes have been added so far.
|
||||
readyNodeCount, _ := e2enode.TotalReady(ctx, f.ClientSet)
|
||||
gomega.Expect(readyNodeCount).To(gomega.Equal(nodeCount))
|
||||
|
||||
// Start a number of schedulable pods to ensure CA reacts.
|
||||
additionalNodes := maxNodes - nodeCount
|
||||
replicas := additionalNodes * replicasPerNode
|
||||
totalMemory := additionalNodes * perNodeReservation
|
||||
rcConfig := reserveMemoryRCConfig(f, "extra-pod", replicas, totalMemory, largeScaleUpTimeout)
|
||||
expectedResult := createClusterPredicates(nodeCount + additionalNodes)
|
||||
config := createScaleUpTestConfig(nodeCount, initialPodReplicas, rcConfig, expectedResult)
|
||||
|
||||
// Test that scale up happens, allowing 1000 unschedulable pods not to be scheduled.
|
||||
testCleanup := simpleScaleUpTestWithTolerance(ctx, f, config, 0, unschedulablePodReplicas)
|
||||
ginkgo.DeferCleanup(testCleanup)
|
||||
})
|
||||
|
||||
})
|
||||
|
||||
func anyKey(input map[string]int) string {
|
||||
for k := range input {
|
||||
return k
|
||||
}
|
||||
return ""
|
||||
}
|
||||
|
||||
func simpleScaleUpTestWithTolerance(ctx context.Context, f *framework.Framework, config *scaleUpTestConfig, tolerateMissingNodeCount int, tolerateMissingPodCount int) func() error {
|
||||
// resize cluster to start size
|
||||
// run rc based on config
|
||||
ginkgo.By(fmt.Sprintf("Running RC %v from config", config.extraPods.Name))
|
||||
start := time.Now()
|
||||
framework.ExpectNoError(e2erc.RunRC(ctx, *config.extraPods))
|
||||
// check results
|
||||
if tolerateMissingNodeCount > 0 {
|
||||
// Tolerate some number of nodes not to be created.
|
||||
minExpectedNodeCount := config.expectedResult.nodes - tolerateMissingNodeCount
|
||||
framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
|
||||
func(size int) bool { return size >= minExpectedNodeCount }, scaleUpTimeout))
|
||||
} else {
|
||||
framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, f.ClientSet, config.expectedResult.nodes, scaleUpTimeout))
|
||||
}
|
||||
klog.Infof("cluster is increased")
|
||||
if tolerateMissingPodCount > 0 {
|
||||
framework.ExpectNoError(waitForCaPodsReadyInNamespace(ctx, f, f.ClientSet, tolerateMissingPodCount))
|
||||
} else {
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, f.ClientSet))
|
||||
}
|
||||
timeTrack(start, fmt.Sprintf("Scale up to %v", config.expectedResult.nodes))
|
||||
return func() error {
|
||||
return e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, config.extraPods.Name)
|
||||
}
|
||||
}
|
||||
|
||||
func simpleScaleUpTest(ctx context.Context, f *framework.Framework, config *scaleUpTestConfig) func() error {
|
||||
return simpleScaleUpTestWithTolerance(ctx, f, config, 0, 0)
|
||||
}
|
||||
|
||||
func reserveMemoryRCConfig(f *framework.Framework, id string, replicas, megabytes int, timeout time.Duration) *testutils.RCConfig {
|
||||
return &testutils.RCConfig{
|
||||
Client: f.ClientSet,
|
||||
Name: id,
|
||||
Namespace: f.Namespace.Name,
|
||||
Timeout: timeout,
|
||||
Image: imageutils.GetPauseImageName(),
|
||||
Replicas: replicas,
|
||||
MemRequest: int64(1024 * 1024 * megabytes / replicas),
|
||||
}
|
||||
}
|
||||
|
||||
func createScaleUpTestConfig(nodes, pods int, extraPods *testutils.RCConfig, expectedResult *clusterPredicates) *scaleUpTestConfig {
|
||||
return &scaleUpTestConfig{
|
||||
initialNodes: nodes,
|
||||
initialPods: pods,
|
||||
extraPods: extraPods,
|
||||
expectedResult: expectedResult,
|
||||
}
|
||||
}
|
||||
|
||||
func createClusterPredicates(nodes int) *clusterPredicates {
|
||||
return &clusterPredicates{
|
||||
nodes: nodes,
|
||||
}
|
||||
}
|
||||
|
||||
func addAnnotation(ctx context.Context, f *framework.Framework, nodes []v1.Node, key, value string) error {
|
||||
for _, node := range nodes {
|
||||
oldData, err := json.Marshal(node)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
|
||||
if node.Annotations == nil {
|
||||
node.Annotations = make(map[string]string)
|
||||
}
|
||||
node.Annotations[key] = value
|
||||
|
||||
newData, err := json.Marshal(node)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
|
||||
patchBytes, err := strategicpatch.CreateTwoWayMergePatch(oldData, newData, v1.Node{})
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
|
||||
_, err = f.ClientSet.CoreV1().Nodes().Patch(ctx, string(node.Name), types.StrategicMergePatchType, patchBytes, metav1.PatchOptions{})
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func createHostPortPodsWithMemory(ctx context.Context, f *framework.Framework, id string, replicas, port, megabytes int, timeout time.Duration) func() error {
|
||||
ginkgo.By(fmt.Sprintf("Running RC which reserves host port and memory"))
|
||||
request := int64(1024 * 1024 * megabytes / replicas)
|
||||
config := &testutils.RCConfig{
|
||||
Client: f.ClientSet,
|
||||
Name: id,
|
||||
Namespace: f.Namespace.Name,
|
||||
Timeout: timeout,
|
||||
Image: imageutils.GetPauseImageName(),
|
||||
Replicas: replicas,
|
||||
HostPorts: map[string]int{"port1": port},
|
||||
MemRequest: request,
|
||||
}
|
||||
err := e2erc.RunRC(ctx, *config)
|
||||
framework.ExpectNoError(err)
|
||||
return func() error {
|
||||
return e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, id)
|
||||
}
|
||||
}
|
||||
|
||||
type podBatch struct {
|
||||
numNodes int
|
||||
podsPerNode int
|
||||
}
|
||||
|
||||
// distributeLoad distributes the pods in the way described by podDostribution,
|
||||
// assuming all pods will have the same memory reservation and all nodes the same
|
||||
// memory capacity. This allows us generate the load on the cluster in the exact
|
||||
// way that we want.
|
||||
//
|
||||
// To achieve this we do the following:
|
||||
// 1. Create replication controllers that eat up all the space that should be
|
||||
// empty after setup, making sure they end up on different nodes by specifying
|
||||
// conflicting host port
|
||||
// 2. Create target RC that will generate the load on the cluster
|
||||
// 3. Remove the rcs created in 1.
|
||||
func distributeLoad(ctx context.Context, f *framework.Framework, namespace string, id string, podDistribution []podBatch,
|
||||
podMemRequestMegabytes int, nodeMemCapacity int, labels map[string]string, timeout time.Duration) {
|
||||
port := 8013
|
||||
// Create load-distribution RCs with one pod per node, reserving all remaining
|
||||
// memory to force the distribution of pods for the target RCs.
|
||||
// The load-distribution RCs will be deleted on function return.
|
||||
totalPods := 0
|
||||
for i, podBatch := range podDistribution {
|
||||
totalPods += podBatch.numNodes * podBatch.podsPerNode
|
||||
remainingMem := nodeMemCapacity - podBatch.podsPerNode*podMemRequestMegabytes
|
||||
replicas := podBatch.numNodes
|
||||
cleanup := createHostPortPodsWithMemory(ctx, f, fmt.Sprintf("load-distribution%d", i), replicas, port, remainingMem*replicas, timeout)
|
||||
defer cleanup()
|
||||
}
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, f.ClientSet))
|
||||
// Create the target RC
|
||||
rcConfig := reserveMemoryRCConfig(f, id, totalPods, totalPods*podMemRequestMegabytes, timeout)
|
||||
framework.ExpectNoError(e2erc.RunRC(ctx, *rcConfig))
|
||||
framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, f.ClientSet))
|
||||
ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, id)
|
||||
}
|
||||
|
||||
func timeTrack(start time.Time, name string) {
|
||||
elapsed := time.Since(start)
|
||||
klog.Infof("%s took %s", name, elapsed)
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,425 +0,0 @@
|
||||
/*
|
||||
Copyright 2016 The Kubernetes Authors.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
*/
|
||||
|
||||
package autoscaling
|
||||
|
||||
import (
|
||||
"context"
|
||||
"fmt"
|
||||
"math"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
v1 "k8s.io/api/core/v1"
|
||||
"k8s.io/apimachinery/pkg/api/resource"
|
||||
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
|
||||
"k8s.io/apimachinery/pkg/labels"
|
||||
"k8s.io/apimachinery/pkg/util/wait"
|
||||
clientset "k8s.io/client-go/kubernetes"
|
||||
"k8s.io/kubernetes/test/e2e/framework"
|
||||
e2enode "k8s.io/kubernetes/test/e2e/framework/node"
|
||||
e2epod "k8s.io/kubernetes/test/e2e/framework/pod"
|
||||
e2eskipper "k8s.io/kubernetes/test/e2e/framework/skipper"
|
||||
admissionapi "k8s.io/pod-security-admission/api"
|
||||
|
||||
"github.com/onsi/ginkgo/v2"
|
||||
)
|
||||
|
||||
// This test requires coredns to be installed on the cluster with autoscaling enabled.
|
||||
// Compare your coredns manifest against the command below
|
||||
// helm template coredns -n kube-system coredns/coredns --set k8sAppLabelOverride=kube-dns --set fullnameOverride=coredns --set autoscaler.enabled=true
|
||||
|
||||
// Constants used in dns-autoscaling test.
|
||||
const (
|
||||
DNSdefaultTimeout = 5 * time.Minute
|
||||
ClusterAddonLabelKey = "k8s-app"
|
||||
DNSLabelName = "kube-dns"
|
||||
)
|
||||
|
||||
var _ = SIGDescribe("DNS horizontal autoscaling", func() {
|
||||
f := framework.NewDefaultFramework("dns-autoscaling")
|
||||
f.NamespacePodSecurityLevel = admissionapi.LevelPrivileged
|
||||
var c clientset.Interface
|
||||
var previousParams map[string]string
|
||||
var configMapNames map[string]string
|
||||
var originDNSReplicasCount int
|
||||
var DNSParams1 DNSParamsLinear
|
||||
var DNSParams2 DNSParamsLinear
|
||||
var DNSParams3 DNSParamsLinear
|
||||
|
||||
ginkgo.BeforeEach(func(ctx context.Context) {
|
||||
e2eskipper.SkipUnlessProviderIs("gce", "gke")
|
||||
c = f.ClientSet
|
||||
|
||||
nodes, err := e2enode.GetReadySchedulableNodes(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
nodeCount := len(nodes.Items)
|
||||
|
||||
ginkgo.By("Collecting original replicas count and DNS scaling params")
|
||||
|
||||
// Check if we are running coredns or kube-dns, the only difference is the name of the autoscaling CM.
|
||||
// The test should be have identically on both dns providers
|
||||
provider, err := detectDNSProvider(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
originDNSReplicasCount, err = getDNSReplicas(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
configMapNames = map[string]string{
|
||||
"kube-dns": "kube-dns-autoscaler",
|
||||
"coredns": "coredns-autoscaler",
|
||||
}
|
||||
|
||||
pcm, err := fetchDNSScalingConfigMap(ctx, c, configMapNames[provider])
|
||||
framework.Logf("original DNS scaling params: %v", pcm)
|
||||
framework.ExpectNoError(err)
|
||||
previousParams = pcm.Data
|
||||
|
||||
if nodeCount <= 500 {
|
||||
DNSParams1 = DNSParamsLinear{
|
||||
nodesPerReplica: 1,
|
||||
}
|
||||
DNSParams2 = DNSParamsLinear{
|
||||
nodesPerReplica: 2,
|
||||
}
|
||||
DNSParams3 = DNSParamsLinear{
|
||||
nodesPerReplica: 3,
|
||||
coresPerReplica: 3,
|
||||
}
|
||||
} else {
|
||||
// In large clusters, avoid creating/deleting too many DNS pods,
|
||||
// it is supposed to be correctness test, not performance one.
|
||||
// The default setup is: 256 cores/replica, 16 nodes/replica.
|
||||
// With nodeCount > 500, nodes/13, nodes/14, nodes/15 and nodes/16
|
||||
// are different numbers.
|
||||
DNSParams1 = DNSParamsLinear{
|
||||
nodesPerReplica: 13,
|
||||
}
|
||||
DNSParams2 = DNSParamsLinear{
|
||||
nodesPerReplica: 14,
|
||||
}
|
||||
DNSParams3 = DNSParamsLinear{
|
||||
nodesPerReplica: 15,
|
||||
coresPerReplica: 15,
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
// This test is separated because it is slow and need to run serially.
|
||||
// Will take around 5 minutes to run on a 4 nodes cluster.
|
||||
// TODO(upodroid) This test will be removed in 1.33 when kubeup is removed
|
||||
// TODO: make it cloud provider agnostic or move it to cloud-provider-gcp repository
|
||||
f.It(f.WithSerial(), f.WithSlow(), f.WithLabel("KubeUp"), f.WithLabel("sig-cloud-provider-gcp"), "kube-dns-autoscaler should scale kube-dns pods when cluster size changed", func(ctx context.Context) {
|
||||
numNodes, err := e2enode.TotalRegistered(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
configMapNames = map[string]string{
|
||||
"kube-dns": "kube-dns-autoscaler",
|
||||
"coredns": "coredns-autoscaler",
|
||||
}
|
||||
provider, err := detectDNSProvider(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Replace the dns autoscaling parameters with testing parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams1)))
|
||||
framework.ExpectNoError(err)
|
||||
defer func() {
|
||||
ginkgo.By("Restoring initial dns autoscaling parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], previousParams))
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Wait for number of running and ready kube-dns pods recover")
|
||||
label := labels.SelectorFromSet(labels.Set(map[string]string{ClusterAddonLabelKey: DNSLabelName}))
|
||||
_, err := e2epod.WaitForPodsWithLabelRunningReady(ctx, c, metav1.NamespaceSystem, label, originDNSReplicasCount, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
}()
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
getExpectReplicasLinear := getExpectReplicasFuncLinear(ctx, c, &DNSParams1)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
originalSizes := make(map[string]int)
|
||||
for _, mig := range strings.Split(framework.TestContext.CloudConfig.NodeInstanceGroup, ",") {
|
||||
size, err := framework.GroupSize(mig)
|
||||
framework.ExpectNoError(err)
|
||||
ginkgo.By(fmt.Sprintf("Initial size of %s: %d", mig, size))
|
||||
originalSizes[mig] = size
|
||||
}
|
||||
|
||||
ginkgo.By("Manually increase cluster size")
|
||||
increasedSizes := make(map[string]int)
|
||||
for key, val := range originalSizes {
|
||||
increasedSizes[key] = val + 1
|
||||
}
|
||||
setMigSizes(increasedSizes)
|
||||
err = WaitForClusterSizeFunc(ctx, c,
|
||||
func(size int) bool { return size == numNodes+len(originalSizes) }, scaleUpTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
getExpectReplicasLinear = getExpectReplicasFuncLinear(ctx, c, &DNSParams1)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Replace the dns autoscaling parameters with another testing parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams3)))
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
getExpectReplicasLinear = getExpectReplicasFuncLinear(ctx, c, &DNSParams3)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Restoring cluster size")
|
||||
setMigSizes(originalSizes)
|
||||
err = e2enode.WaitForReadyNodes(ctx, c, numNodes, scaleDownTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
})
|
||||
|
||||
ginkgo.It("kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios", func(ctx context.Context) {
|
||||
|
||||
configMapNames = map[string]string{
|
||||
"kube-dns": "kube-dns-autoscaler",
|
||||
"coredns": "coredns-autoscaler",
|
||||
}
|
||||
provider, err := detectDNSProvider(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Replace the dns autoscaling parameters with testing parameters")
|
||||
cm := packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams1))
|
||||
framework.Logf("Updating the following cm: %v", cm)
|
||||
err = updateDNSScalingConfigMap(ctx, c, cm)
|
||||
framework.ExpectNoError(err)
|
||||
defer func() {
|
||||
ginkgo.By("Restoring initial dns autoscaling parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], previousParams))
|
||||
framework.ExpectNoError(err)
|
||||
}()
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
getExpectReplicasLinear := getExpectReplicasFuncLinear(ctx, c, &DNSParams1)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("--- Scenario: should scale kube-dns based on changed parameters ---")
|
||||
ginkgo.By("Replace the dns autoscaling parameters with another testing parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams3)))
|
||||
framework.ExpectNoError(err)
|
||||
ginkgo.By("Wait for kube-dns scaled to expected number")
|
||||
getExpectReplicasLinear = getExpectReplicasFuncLinear(ctx, c, &DNSParams3)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("--- Scenario: should re-create scaling parameters with default value when parameters got deleted ---")
|
||||
ginkgo.By("Delete the ConfigMap for autoscaler")
|
||||
err = deleteDNSScalingConfigMap(ctx, c, configMapNames[provider])
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Wait for the ConfigMap got re-created")
|
||||
_, err = waitForDNSConfigMapCreated(ctx, c, DNSdefaultTimeout, configMapNames[provider])
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Replace the dns autoscaling parameters with another testing parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams2)))
|
||||
framework.ExpectNoError(err)
|
||||
ginkgo.By("Wait for kube-dns/coredns scaled to expected number")
|
||||
getExpectReplicasLinear = getExpectReplicasFuncLinear(ctx, c, &DNSParams2)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("--- Scenario: should recover after autoscaler pod got deleted ---")
|
||||
ginkgo.By("Delete the autoscaler pod for kube-dns/coredns")
|
||||
err = deleteDNSAutoscalerPod(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
|
||||
ginkgo.By("Replace the dns autoscaling parameters with another testing parameters")
|
||||
err = updateDNSScalingConfigMap(ctx, c, packDNSScalingConfigMap(configMapNames[provider], packLinearParams(&DNSParams1)))
|
||||
framework.ExpectNoError(err)
|
||||
ginkgo.By("Wait for kube-dns/coredns scaled to expected number")
|
||||
getExpectReplicasLinear = getExpectReplicasFuncLinear(ctx, c, &DNSParams1)
|
||||
err = waitForDNSReplicasSatisfied(ctx, c, getExpectReplicasLinear, DNSdefaultTimeout)
|
||||
framework.ExpectNoError(err)
|
||||
})
|
||||
})
|
||||
|
||||
// DNSParamsLinear is a struct for number of DNS pods.
|
||||
type DNSParamsLinear struct {
|
||||
nodesPerReplica float64
|
||||
coresPerReplica float64
|
||||
min int
|
||||
max int
|
||||
}
|
||||
|
||||
type getExpectReplicasFunc func(c clientset.Interface) int
|
||||
|
||||
func getExpectReplicasFuncLinear(ctx context.Context, c clientset.Interface, params *DNSParamsLinear) getExpectReplicasFunc {
|
||||
return func(c clientset.Interface) int {
|
||||
var replicasFromNodes float64
|
||||
var replicasFromCores float64
|
||||
nodes, err := e2enode.GetReadyNodesIncludingTainted(ctx, c)
|
||||
framework.ExpectNoError(err)
|
||||
if params.nodesPerReplica > 0 {
|
||||
replicasFromNodes = math.Ceil(float64(len(nodes.Items)) / params.nodesPerReplica)
|
||||
}
|
||||
if params.coresPerReplica > 0 {
|
||||
replicasFromCores = math.Ceil(float64(getSchedulableCores(nodes.Items)) / params.coresPerReplica)
|
||||
}
|
||||
return int(math.Max(1.0, math.Max(replicasFromNodes, replicasFromCores)))
|
||||
}
|
||||
}
|
||||
|
||||
func getSchedulableCores(nodes []v1.Node) int64 {
|
||||
var sc resource.Quantity
|
||||
for _, node := range nodes {
|
||||
if !node.Spec.Unschedulable {
|
||||
sc.Add(node.Status.Allocatable[v1.ResourceCPU])
|
||||
}
|
||||
}
|
||||
return sc.Value()
|
||||
}
|
||||
|
||||
func detectDNSProvider(ctx context.Context, c clientset.Interface) (string, error) {
|
||||
cm, err := c.CoreV1().ConfigMaps(metav1.NamespaceSystem).Get(ctx, "coredns-autoscaler", metav1.GetOptions{})
|
||||
if cm != nil && err == nil {
|
||||
return "coredns", nil
|
||||
}
|
||||
|
||||
cm, err = c.CoreV1().ConfigMaps(metav1.NamespaceSystem).Get(ctx, "kube-dns-autoscaler", metav1.GetOptions{})
|
||||
if cm != nil && err == nil {
|
||||
return "kube-dns", nil
|
||||
}
|
||||
|
||||
return "", fmt.Errorf("the cluster doesn't have kube-dns or coredns autoscaling configured")
|
||||
}
|
||||
|
||||
func fetchDNSScalingConfigMap(ctx context.Context, c clientset.Interface, configMapName string) (*v1.ConfigMap, error) {
|
||||
cm, err := c.CoreV1().ConfigMaps(metav1.NamespaceSystem).Get(ctx, configMapName, metav1.GetOptions{})
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
return cm, nil
|
||||
}
|
||||
|
||||
func deleteDNSScalingConfigMap(ctx context.Context, c clientset.Interface, configMapName string) error {
|
||||
if err := c.CoreV1().ConfigMaps(metav1.NamespaceSystem).Delete(ctx, configMapName, metav1.DeleteOptions{}); err != nil {
|
||||
return err
|
||||
}
|
||||
framework.Logf("DNS autoscaling ConfigMap deleted.")
|
||||
return nil
|
||||
}
|
||||
|
||||
func packLinearParams(params *DNSParamsLinear) map[string]string {
|
||||
paramsMap := make(map[string]string)
|
||||
paramsMap["linear"] = fmt.Sprintf("{\"nodesPerReplica\": %v,\"coresPerReplica\": %v,\"min\": %v,\"max\": %v}",
|
||||
params.nodesPerReplica,
|
||||
params.coresPerReplica,
|
||||
params.min,
|
||||
params.max)
|
||||
return paramsMap
|
||||
}
|
||||
|
||||
func packDNSScalingConfigMap(configMapName string, params map[string]string) *v1.ConfigMap {
|
||||
configMap := v1.ConfigMap{}
|
||||
configMap.ObjectMeta.Name = configMapName
|
||||
configMap.ObjectMeta.Namespace = metav1.NamespaceSystem
|
||||
configMap.Data = params
|
||||
return &configMap
|
||||
}
|
||||
|
||||
func updateDNSScalingConfigMap(ctx context.Context, c clientset.Interface, configMap *v1.ConfigMap) error {
|
||||
_, err := c.CoreV1().ConfigMaps(metav1.NamespaceSystem).Update(ctx, configMap, metav1.UpdateOptions{})
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
framework.Logf("DNS autoscaling ConfigMap updated.")
|
||||
return nil
|
||||
}
|
||||
|
||||
func getDNSReplicas(ctx context.Context, c clientset.Interface) (int, error) {
|
||||
label := labels.SelectorFromSet(labels.Set(map[string]string{ClusterAddonLabelKey: DNSLabelName}))
|
||||
listOpts := metav1.ListOptions{LabelSelector: label.String()}
|
||||
deployments, err := c.AppsV1().Deployments(metav1.NamespaceSystem).List(ctx, listOpts)
|
||||
if err != nil {
|
||||
return 0, err
|
||||
}
|
||||
if len(deployments.Items) != 1 {
|
||||
return 0, fmt.Errorf("expected 1 DNS deployment, got %v", len(deployments.Items))
|
||||
}
|
||||
|
||||
deployment := deployments.Items[0]
|
||||
return int(*(deployment.Spec.Replicas)), nil
|
||||
}
|
||||
|
||||
func deleteDNSAutoscalerPod(ctx context.Context, c clientset.Interface) error {
|
||||
selector, _ := labels.Parse(fmt.Sprintf("%s in (kube-dns-autoscaler, coredns-autoscaler)", ClusterAddonLabelKey))
|
||||
listOpts := metav1.ListOptions{LabelSelector: selector.String()}
|
||||
pods, err := c.CoreV1().Pods(metav1.NamespaceSystem).List(ctx, listOpts)
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
if len(pods.Items) != 1 {
|
||||
return fmt.Errorf("expected 1 autoscaler pod, got %v", len(pods.Items))
|
||||
}
|
||||
|
||||
podName := pods.Items[0].Name
|
||||
if err := c.CoreV1().Pods(metav1.NamespaceSystem).Delete(ctx, podName, metav1.DeleteOptions{}); err != nil {
|
||||
return err
|
||||
}
|
||||
framework.Logf("DNS autoscaling pod %v deleted.", podName)
|
||||
return nil
|
||||
}
|
||||
|
||||
func waitForDNSReplicasSatisfied(ctx context.Context, c clientset.Interface, getExpected getExpectReplicasFunc, timeout time.Duration) (err error) {
|
||||
var current int
|
||||
var expected int
|
||||
framework.Logf("Waiting up to %v for kube-dns to reach expected replicas", timeout)
|
||||
condition := func(ctx context.Context) (bool, error) {
|
||||
current, err = getDNSReplicas(ctx, c)
|
||||
if err != nil {
|
||||
return false, err
|
||||
}
|
||||
expected = getExpected(c)
|
||||
if current != expected {
|
||||
framework.Logf("Replicas not as expected: got %v, expected %v", current, expected)
|
||||
return false, nil
|
||||
}
|
||||
return true, nil
|
||||
}
|
||||
|
||||
if err = wait.PollUntilContextTimeout(ctx, 2*time.Second, timeout, false, condition); err != nil {
|
||||
return fmt.Errorf("err waiting for DNS replicas to satisfy %v, got %v: %w", expected, current, err)
|
||||
}
|
||||
framework.Logf("kube-dns reaches expected replicas: %v", expected)
|
||||
return nil
|
||||
}
|
||||
|
||||
func waitForDNSConfigMapCreated(ctx context.Context, c clientset.Interface, timeout time.Duration, configMapName string) (configMap *v1.ConfigMap, err error) {
|
||||
framework.Logf("Waiting up to %v for DNS autoscaling ConfigMap to be re-created", timeout)
|
||||
condition := func(ctx context.Context) (bool, error) {
|
||||
configMap, err = fetchDNSScalingConfigMap(ctx, c, configMapName)
|
||||
if err != nil {
|
||||
return false, nil
|
||||
}
|
||||
return true, nil
|
||||
}
|
||||
|
||||
if err = wait.PollUntilContextTimeout(ctx, time.Second, timeout, false, condition); err != nil {
|
||||
return nil, fmt.Errorf("err waiting for DNS autoscaling ConfigMap got re-created: %w", err)
|
||||
}
|
||||
return configMap, nil
|
||||
}
|
||||
@@ -20,9 +20,4 @@ const (
|
||||
// NVIDIAGPUResourceName is the extended name of the GPU resource since v1.8
|
||||
// this uses the device plugin mechanism
|
||||
NVIDIAGPUResourceName = "nvidia.com/gpu"
|
||||
|
||||
// GPUDevicePluginDSYAML is the official Google Device Plugin Daemonset NVIDIA GPU manifest for GKE
|
||||
// TODO: Parametrize it by making it a feature in TestFramework.
|
||||
// so we can override the daemonset in other setups (non COS).
|
||||
GPUDevicePluginDSYAML = "https://raw.githubusercontent.com/kubernetes/kubernetes/master/cluster/addons/device-plugins/nvidia-gpu/daemonset.yaml"
|
||||
)
|
||||
|
||||
@@ -32,7 +32,6 @@ import (
|
||||
internalapi "k8s.io/cri-api/pkg/apis"
|
||||
runtimeapi "k8s.io/cri-api/pkg/apis/runtime/v1"
|
||||
commontest "k8s.io/kubernetes/test/e2e/common"
|
||||
e2egpu "k8s.io/kubernetes/test/e2e/framework/gpu"
|
||||
e2emanifest "k8s.io/kubernetes/test/e2e/framework/manifest"
|
||||
e2epod "k8s.io/kubernetes/test/e2e/framework/pod"
|
||||
e2etestfiles "k8s.io/kubernetes/test/e2e/framework/testfiles"
|
||||
@@ -83,11 +82,6 @@ func updateImageAllowList(ctx context.Context) {
|
||||
} else {
|
||||
e2epod.ImagePrePullList.Insert(sriovDevicePluginImage)
|
||||
}
|
||||
if gpuDevicePluginImage, err := getGPUDevicePluginImage(ctx); err != nil {
|
||||
klog.Errorln(err)
|
||||
} else {
|
||||
e2epod.ImagePrePullList.Insert(gpuDevicePluginImage)
|
||||
}
|
||||
if samplePluginImage, err := getContainerImageFromE2ETestDaemonset(SampleDevicePluginDSYAML); err != nil {
|
||||
klog.Errorln(err)
|
||||
} else {
|
||||
@@ -217,21 +211,6 @@ func PrePullAllImages() error {
|
||||
return utilerrors.NewAggregate(pullErrs)
|
||||
}
|
||||
|
||||
// getGPUDevicePluginImage returns the image of GPU device plugin.
|
||||
func getGPUDevicePluginImage(ctx context.Context) (string, error) {
|
||||
ds, err := e2emanifest.DaemonSetFromURL(ctx, e2egpu.GPUDevicePluginDSYAML)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("failed to parse the device plugin image: %w", err)
|
||||
}
|
||||
if ds == nil {
|
||||
return "", fmt.Errorf("failed to parse the device plugin image: the extracted DaemonSet is nil")
|
||||
}
|
||||
if len(ds.Spec.Template.Spec.Containers) < 1 {
|
||||
return "", fmt.Errorf("failed to parse the device plugin image: cannot extract the container from YAML")
|
||||
}
|
||||
return ds.Spec.Template.Spec.Containers[0].Image, nil
|
||||
}
|
||||
|
||||
func getContainerImageFromE2ETestDaemonset(dsYamlPath string) (string, error) {
|
||||
data, err := e2etestfiles.Read(dsYamlPath)
|
||||
if err != nil {
|
||||
|
||||
@@ -1,27 +0,0 @@
|
||||
#cloud-config
|
||||
|
||||
runcmd:
|
||||
- modprobe configs
|
||||
# Install GPU drivers - https://cloud.google.com/container-optimized-os/docs/how-to/run-gpus
|
||||
- cos-extensions install gpu
|
||||
- mount --bind /var/lib/nvidia /var/lib/nvidia
|
||||
- mount -o remount,exec /var/lib/nvidia /var/lib/nvidia
|
||||
# Run nvidia-smi to verify installation
|
||||
- /var/lib/nvidia/bin/nvidia-smi
|
||||
# Remove build containers. They're very large.
|
||||
- docker rm -f $(docker ps -aq)
|
||||
# Standard installation proceeds
|
||||
- mount /tmp /tmp -o remount,exec,suid
|
||||
- usermod -a -G docker jenkins
|
||||
- mkdir -p /var/lib/kubelet
|
||||
- mkdir -p /home/kubernetes/containerized_mounter/rootfs
|
||||
- mount --bind /home/kubernetes/containerized_mounter/ /home/kubernetes/containerized_mounter/
|
||||
- mount -o remount, exec /home/kubernetes/containerized_mounter/
|
||||
- wget https://storage.googleapis.com/kubernetes-release/gci-mounter/mounter.tar -O /tmp/mounter.tar
|
||||
- tar xvf /tmp/mounter.tar -C /home/kubernetes/containerized_mounter/rootfs
|
||||
- mkdir -p /home/kubernetes/containerized_mounter/rootfs/var/lib/kubelet
|
||||
- mount --rbind /var/lib/kubelet /home/kubernetes/containerized_mounter/rootfs/var/lib/kubelet
|
||||
- mount --make-rshared /home/kubernetes/containerized_mounter/rootfs/var/lib/kubelet
|
||||
- mount --bind /proc /home/kubernetes/containerized_mounter/rootfs/proc
|
||||
- mount --bind /dev /home/kubernetes/containerized_mounter/rootfs/dev
|
||||
- rm /tmp/mounter.tar
|
||||
Reference in New Issue
Block a user