Added device plugin e2e kubelet failure test

Signed-off-by: Renaud Gaubert <renaud.gaubert@gmail.com>

test/e2e/framework/BUILD

@@ -17,6 +17,7 @@ go_library(
         "framework.go",
         "get-kubemark-resource-usage.go",
         "google_compute.go",
+        "gpu_util.go",
         "ingress_utils.go",
         "jobs_util.go",
         "kubelet_stats.go",
@@ -121,6 +122,7 @@ go_library(
         "//vendor/k8s.io/apimachinery/pkg/util/sets:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/util/uuid:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/util/wait:go_default_library",
+        "//vendor/k8s.io/apimachinery/pkg/util/yaml:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/version:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/watch:go_default_library",
         "//vendor/k8s.io/client-go/discovery:go_default_library",

test/e2e/framework/gpu_util.go

@@ -0,0 +1,66 @@
+/*
+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 framework
+
+import (
+	"k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/uuid"
+)
+
+const (
+	// GPUResourceName is the extended name of the GPU resource since v1.8
+	// this uses the device plugin mechanism
+	NVIDIAGPUResourceName = "nvidia.com/gpu"
+
+	// TODO: Parametrize it by making it a feature in TestFramework.
+	// so we can override the daemonset in other setups (non COS).
+	// GPUDevicePluginDSYAML is the official Google Device Plugin Daemonset NVIDIA GPU manifest for GKE
+	GPUDevicePluginDSYAML = "https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/master/device-plugin-daemonset.yaml"
+)
+
+// TODO make this generic and not linked to COS only
+// NumberOfGPUs returs the number of GPUs advertised by a node
+// This is based on the Device Plugin system and expected to run on a COS based node
+// After the NVIDIA drivers were installed
+func NumberOfNVIDIAGPUs(node *v1.Node) int64 {
+	val, ok := node.Status.Capacity[NVIDIAGPUResourceName]
+
+	if !ok {
+		return 0
+	}
+
+	return val.Value()
+}
+
+// NVIDIADevicePlugin returns the official Google Device Plugin pod for NVIDIA GPU in GKE
+func NVIDIADevicePlugin(ns string) *v1.Pod {
+	ds := DsFromManifest(GPUDevicePluginDSYAML)
+	p := &v1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "device-plugin-nvidia-gpu-" + string(uuid.NewUUID()),
+			Namespace: ns,
+		},
+
+		Spec: ds.Spec.Template.Spec,
+	}
+
+	// Remove NVIDIA drivers installation
+	p.Spec.InitContainers = []v1.Container{}
+
+	return p
+}

test/e2e/framework/util.go

@@ -65,6 +65,7 @@ import (
 	"k8s.io/apimachinery/pkg/util/sets"
 	"k8s.io/apimachinery/pkg/util/uuid"
 	"k8s.io/apimachinery/pkg/util/wait"
+	utilyaml "k8s.io/apimachinery/pkg/util/yaml"
 	"k8s.io/apimachinery/pkg/watch"
 	"k8s.io/client-go/discovery"
 	"k8s.io/client-go/dynamic"
@@ -5015,3 +5016,33 @@ func DumpDebugInfo(c clientset.Interface, ns string) {
 func IsRetryableAPIError(err error) bool {
 	return apierrs.IsTimeout(err) || apierrs.IsServerTimeout(err) || apierrs.IsTooManyRequests(err) || apierrs.IsInternalError(err)
 }
+
+// DsFromManifest reads a .json/yaml file and returns the daemonset in it.
+func DsFromManifest(url string) *extensions.DaemonSet {
+	var controller extensions.DaemonSet
+	Logf("Parsing ds from %v", url)
+
+	var response *http.Response
+	var err error
+
+	for i := 1; i <= 5; i++ {
+		response, err = http.Get(url)
+		if err == nil && response.StatusCode == 200 {
+			break
+		}
+		time.Sleep(time.Duration(i) * time.Second)
+	}
+
+	Expect(err).NotTo(HaveOccurred())
+	Expect(response.StatusCode).To(Equal(200))
+	defer response.Body.Close()
+
+	data, err := ioutil.ReadAll(response.Body)
+	Expect(err).NotTo(HaveOccurred())
+
+	json, err := utilyaml.ToJSON(data)
+	Expect(err).NotTo(HaveOccurred())
+
+	Expect(runtime.DecodeInto(api.Codecs.UniversalDecoder(), json, &controller)).NotTo(HaveOccurred())
+	return &controller
+}

test/e2e/scheduling/BUILD

@@ -34,17 +34,14 @@ go_library(
         "//vendor/github.com/onsi/gomega:go_default_library",
         "//vendor/github.com/stretchr/testify/assert:go_default_library",
         "//vendor/k8s.io/api/core/v1:go_default_library",
-        "//vendor/k8s.io/api/extensions/v1beta1:go_default_library",
         "//vendor/k8s.io/api/scheduling/v1alpha1:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/api/errors:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/api/resource:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/apis/meta/v1:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/labels:go_default_library",
-        "//vendor/k8s.io/apimachinery/pkg/runtime:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/types:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/util/sets:go_default_library",
         "//vendor/k8s.io/apimachinery/pkg/util/uuid:go_default_library",
-        "//vendor/k8s.io/apimachinery/pkg/util/yaml:go_default_library",
         "//vendor/k8s.io/client-go/kubernetes:go_default_library",
     ],
 )

test/e2e/scheduling/nvidia-gpus.go

@@ -17,19 +17,13 @@ limitations under the License.
 package scheduling
 
 import (
-	"io/ioutil"
-	"net/http"
 	"strings"
 	"time"
 
 	"k8s.io/api/core/v1"
-	extensions "k8s.io/api/extensions/v1beta1"
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
-	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/util/uuid"
-	utilyaml "k8s.io/apimachinery/pkg/util/yaml"
-	"k8s.io/kubernetes/pkg/api"
 	"k8s.io/kubernetes/test/e2e/framework"
 	imageutils "k8s.io/kubernetes/test/utils/image"
 
@@ -168,8 +162,8 @@ func testNvidiaGPUsOnCOS(f *framework.Framework) {
 	framework.Logf("Cluster is running on COS. Proceeding with test")
 
 	if f.BaseName == "device-plugin-gpus" {
-		dsYamlUrl = "https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/master/device-plugin-daemonset.yaml"
-		gpuResourceName = "nvidia.com/gpu"
+		dsYamlUrl = framework.GPUDevicePluginDSYAML
+		gpuResourceName = framework.NVIDIAGPUResourceName
 		podCreationFunc = makeCudaAdditionDevicePluginTestPod
 	} else {
 		dsYamlUrl = "https://raw.githubusercontent.com/ContainerEngine/accelerators/master/cos-nvidia-gpu-installer/daemonset.yaml"
@@ -180,7 +174,7 @@ func testNvidiaGPUsOnCOS(f *framework.Framework) {
 	// GPU drivers might have already been installed.
 	if !areGPUsAvailableOnAllSchedulableNodes(f) {
 		// Install Nvidia Drivers.
-		ds := dsFromManifest(dsYamlUrl)
+		ds := framework.DsFromManifest(dsYamlUrl)
 		ds.Namespace = f.Namespace.Name
 		_, err := f.ClientSet.Extensions().DaemonSets(f.Namespace.Name).Create(ds)
 		framework.ExpectNoError(err, "failed to create daemonset")
@@ -202,34 +196,6 @@ func testNvidiaGPUsOnCOS(f *framework.Framework) {
 	}
 }
 
-// dsFromManifest reads a .json/yaml file and returns the daemonset in it.
-func dsFromManifest(url string) *extensions.DaemonSet {
-	var controller extensions.DaemonSet
-	framework.Logf("Parsing ds from %v", url)
-
-	var response *http.Response
-	var err error
-	for i := 1; i <= 5; i++ {
-		response, err = http.Get(url)
-		if err == nil && response.StatusCode == 200 {
-			break
-		}
-		time.Sleep(time.Duration(i) * time.Second)
-	}
-	Expect(err).NotTo(HaveOccurred())
-	Expect(response.StatusCode).To(Equal(200))
-	defer response.Body.Close()
-
-	data, err := ioutil.ReadAll(response.Body)
-	Expect(err).NotTo(HaveOccurred())
-
-	json, err := utilyaml.ToJSON(data)
-	Expect(err).NotTo(HaveOccurred())
-
-	Expect(runtime.DecodeInto(api.Codecs.UniversalDecoder(), json, &controller)).NotTo(HaveOccurred())
-	return &controller
-}
-
 var _ = SIGDescribe("[Feature:GPU]", func() {
 	f := framework.NewDefaultFramework("gpus")
 	It("run Nvidia GPU tests on Container Optimized OS only", func() {
@@ -247,7 +213,7 @@ var _ = SIGDescribe("[Feature:GPUDevicePlugin]", func() {
 
 		// 2. Verifies that when the device plugin DaemonSet is removed, resource capacity drops to zero.
 		By("Deleting device plugin daemonset")
-		ds := dsFromManifest(dsYamlUrl)
+		ds := framework.DsFromManifest(dsYamlUrl)
 		falseVar := false
 		err := f.ClientSet.Extensions().DaemonSets(f.Namespace.Name).Delete(ds.Name, &metav1.DeleteOptions{OrphanDependents: &falseVar})
 		framework.ExpectNoError(err, "failed to delete daemonset")

test/e2e_node/BUILD

@@ -12,6 +12,7 @@ go_library(
         "container.go",
         "doc.go",
         "docker_util.go",
+        "gpu_device_plugin.go",
         "gpus.go",
         "image_list.go",
         "simple_mount.go",

test/e2e_node/gpu_device_plugin.go

@@ -0,0 +1,164 @@
+/*
+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 e2e_node
+
+import (
+	"fmt"
+	"os/exec"
+	"time"
+
+	"k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/api/resource"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/uuid"
+	"k8s.io/kubernetes/pkg/kubelet/apis/kubeletconfig"
+	"k8s.io/kubernetes/test/e2e/framework"
+
+	. "github.com/onsi/ginkgo"
+	. "github.com/onsi/gomega"
+)
+
+const (
+	devicePluginFeatureGate = "DevicePlugins=true"
+	testPodNamePrefix       = "nvidia-gpu-"
+	sleepTimeout            = 30
+)
+
+// Serial because the test restarts Kubelet
+var _ = framework.KubeDescribe("NVIDIA GPU Device Plugin [Feature:GPUDevicePlugin] [Serial] [Disruptive]", func() {
+	f := framework.NewDefaultFramework("device-plugin-gpus-errors")
+
+	Context("", func() {
+		BeforeEach(func() {
+			By("Ensuring that Nvidia GPUs exists on the node")
+			if !checkIfNvidiaGPUsExistOnNode() {
+				Skip("Nvidia GPUs do not exist on the node. Skipping test.")
+			}
+
+			By("Enabling support for Device Plugin")
+			tempSetCurrentKubeletConfig(f, func(initialConfig *kubeletconfig.KubeletConfiguration) {
+				initialConfig.FeatureGates += "," + devicePluginFeatureGate
+			})
+
+			By("Creating the Google Device Plugin pod for NVIDIA GPU in GKE")
+			f.PodClient().CreateSync(framework.NVIDIADevicePlugin(f.Namespace.Name))
+
+			By("Waiting for GPUs to become available on the local node")
+			Eventually(framework.NumberOfNVIDIAGPUs(getLocalNode(f)) != 0, time.Minute, time.Second).Should(BeTrue())
+
+			if framework.NumberOfNVIDIAGPUs(getLocalNode(f)) < 2 {
+				Skip("Not enough GPUs to execute this test (at least two needed)")
+			}
+		})
+
+		AfterEach(func() {
+			l, err := f.PodClient().List(metav1.ListOptions{})
+			framework.ExpectNoError(err)
+
+			for _, p := range l.Items {
+				if p.Namespace != f.Namespace.Name {
+					continue
+				}
+
+				f.PodClient().Delete(p.Name, &metav1.DeleteOptions{})
+			}
+		})
+
+		It("checks that when Kubelet restarts exclusive GPU assignation to pods is kept.", func() {
+			n := getLocalNode(f)
+
+			By("Creating one GPU pod on a node with at least two GPUs")
+			p1 := f.PodClient().CreateSync(makeCudaPauseImage())
+			cmd := fmt.Sprintf("exec %s %s nvidia-smi -L", n.Name, p1.Spec.Containers[0].Name)
+			uuid1, _ := framework.RunKubectl(cmd)
+
+			By("Restarting Kubelet and waiting for the current running pod to restart")
+			restartKubelet(f)
+			Eventually(func() bool {
+				p, err := f.PodClient().Get(p1.Name, metav1.GetOptions{})
+				framework.ExpectNoError(err)
+
+				return p.Status.ContainerStatuses[0].RestartCount != p1.Status.ContainerStatuses[0].RestartCount
+			}, 2*sleepTimeout)
+
+			By("Confirming that after a kubelet and pod restart, GPU assignement is kept")
+			uuid1Restart, _ := framework.RunKubectl(cmd)
+			Expect(uuid1Restart).To(Equal(uuid1))
+
+			By("Restarting Kubelet and creating another pod")
+			restartKubelet(f)
+			p2 := f.PodClient().CreateSync(makeCudaPauseImage())
+
+			By("Checking that pods got a different GPU")
+			cmd = fmt.Sprintf("exec %s %s nvidia-smi -L", n.Name, p2.Spec.Containers[0].Name)
+			uuid2, _ := framework.RunKubectl(cmd)
+			Expect(uuid1).To(Not(Equal(uuid2)))
+
+			// Cleanup
+			f.PodClient().DeleteSync(p1.Name, &metav1.DeleteOptions{}, framework.DefaultPodDeletionTimeout)
+			f.PodClient().DeleteSync(p2.Name, &metav1.DeleteOptions{}, framework.DefaultPodDeletionTimeout)
+		})
+	})
+})
+
+func makeCudaPauseImage() *v1.Pod {
+	podName := testPodNamePrefix + string(uuid.NewUUID())
+
+	return &v1.Pod{
+		ObjectMeta: metav1.ObjectMeta{Name: podName},
+		Spec: v1.PodSpec{
+			RestartPolicy: v1.RestartPolicyAlways,
+			Containers: []v1.Container{{
+				Name:    "cuda-pause",
+				Image:   "nvidia/cuda",
+				Command: []string{"sleep", string(sleepTimeout)},
+
+				Resources: v1.ResourceRequirements{
+					Limits:   newDecimalResourceList(framework.NVIDIAGPUResourceName, 1),
+					Requests: newDecimalResourceList(framework.NVIDIAGPUResourceName, 1),
+				},
+			}},
+		},
+	}
+}
+
+func newDecimalResourceList(name v1.ResourceName, quantity int64) v1.ResourceList {
+	return v1.ResourceList{name: *resource.NewQuantity(quantity, resource.DecimalSI)}
+}
+
+// TODO: Find a uniform way to deal with systemctl/initctl/service operations. #34494
+func restartKubelet(f *framework.Framework) {
+	stdout1, err1 := exec.Command("sudo", "systemctl", "restart", "kubelet").CombinedOutput()
+	if err1 == nil {
+		return
+	}
+
+	stdout2, err2 := exec.Command("sudo", "/etc/init.d/kubelet", "restart").CombinedOutput()
+	if err2 == nil {
+		return
+	}
+
+	stdout3, err3 := exec.Command("sudo", "service", "kubelet", "restart").CombinedOutput()
+	if err3 == nil {
+		return
+	}
+
+	framework.Failf("Failed to trigger kubelet restart with systemctl/initctl/service operations:"+
+		"\nsystemclt: %v, %v"+
+		"\ninitctl:   %v, %v"+
+		"\nservice:   %v, %v", err1, stdout1, err2, stdout2, err3, stdout3)
+}