feat: maintain the er to support gpu sharing

main.go

@@ -22,17 +22,21 @@ import (
 	"syscall"
 
 	"github.com/NVIDIA/gpu-monitoring-tools/bindings/go/nvml"
+	resourcev1beta1 "github.com/caicloud/clientset/customclient/typed/resource/v1beta1"
+	"github.com/caicloud/clientset/kubernetes"
 	"github.com/fsnotify/fsnotify"
-	pluginapi "k8s.io/kubernetes/pkg/kubelet/apis/deviceplugin/v1beta1"
+	"k8s.io/client-go/rest"
+	pluginapi "k8s.io/kubelet/pkg/apis/deviceplugin/v1beta1"
 )
 
-func getAllPlugins() []*NvidiaDevicePlugin {
+func getAllPlugins(clientset resourcev1beta1.ExtendedResourceInterface) []*NvidiaDevicePlugin {
 	return []*NvidiaDevicePlugin{
 		NewNvidiaDevicePlugin(
+			clientset,
 			"nvidia.com/gpu",
 			NewGpuDeviceManager(),
 			"NVIDIA_VISIBLE_DEVICES",
-			pluginapi.DevicePluginPath + "nvidia.sock"),
+			pluginapi.DevicePluginPath+"nvidia.sock"),
 	}
 }
 
@@ -59,8 +63,10 @@ func main() {
 	log.Println("Starting OS watcher.")
 	sigs := newOSWatcher(syscall.SIGHUP, syscall.SIGINT, syscall.SIGTERM, syscall.SIGQUIT)
 
+	clientset := createClientset()
+
 	log.Println("Retreiving plugins.")
-	plugins := getAllPlugins()
+	plugins := getAllPlugins(clientset)
 
 restart:
 	// Loop through all plugins, idempotently stopping them, and then starting
@@ -72,7 +78,8 @@ restart:
 		p.Stop()
 
 		// Just continue if there are no devices to serve for plugin p.
-		if len(p.Devices()) == 0 {
+		devs, _ := p.Devices()
+		if len(devs) == 0 {
 			continue
 		}
 
@@ -131,3 +138,16 @@ events:
 		}
 	}
 }
+
+func createClientset() resourcev1beta1.ExtendedResourceInterface {
+	config, err := rest.InClusterConfig()
+	if err != nil {
+		log.Panicf("unable to create kubeconfig: %+v", err)
+	}
+
+	clientset, err := kubernetes.NewForConfig(config)
+	if err != nil {
+		log.Panicf("unable to create clientset: %+v", err)
+	}
+	return clientset.Custom().ResourceV1beta1().ExtendedResources()
+}

nvidia-device-plugin-rbac.yaml

@@ -0,0 +1,35 @@
+kind: ClusterRole
+apiVersion: rbac.authorization.k8s.io/v1
+metadata:
+  name: nvidia-device-plugin-cr
+rules:
+  - apiGroups:
+      - "resource.caicloud.io"
+    resources:
+      - extendedresources
+    verbs:
+      - create
+      - get
+      - list
+      - update
+      - delete
+---
+apiVersion: v1
+kind: ServiceAccount
+metadata:
+  name: nvidia-device-plugin-sa
+  namespace: kube-system
+---
+kind: ClusterRoleBinding
+apiVersion: rbac.authorization.k8s.io/v1
+metadata:
+  name: nvidia-device-plugin-crb
+  namespace: kube-system
+roleRef:
+  apiGroup: rbac.authorization.k8s.io
+  kind: ClusterRole
+  name: nvidia-device-plugin-cr
+subjects:
+  - kind: ServiceAccount
+    name: nvidia-device-plugin-sa
+    namespace: kube-system

nvidia-device-plugin.yml

@@ -15,7 +15,7 @@
 apiVersion: apps/v1
 kind: DaemonSet
 metadata:
-  name: nvidia-device-plugin-daemonset
+  name: nvidia-device-plugin-caicloud
   namespace: kube-system
 spec:
   selector:
@@ -25,29 +25,23 @@ spec:
     type: RollingUpdate
   template:
     metadata:
-      # This annotation is deprecated. Kept here for backward compatibility
-      # See https://kubernetes.io/docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods/
-      annotations:
-        scheduler.alpha.kubernetes.io/critical-pod: ""
       labels:
         name: nvidia-device-plugin-ds
     spec:
-      tolerations:
-      # This toleration is deprecated. Kept here for backward compatibility
-      # See https://kubernetes.io/docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods/
-      - key: CriticalAddonsOnly
-        operator: Exists
-      - key: nvidia.com/gpu
-        operator: Exists
-        effect: NoSchedule
+      serviceAccount: nvidia-device-plugin-sa
       # Mark this pod as a critical add-on; when enabled, the critical add-on
       # scheduler reserves resources for critical add-on pods so that they can
       # be rescheduled after a failure.
       # See https://kubernetes.io/docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods/
       priorityClassName: "system-node-critical"
       containers:
-      - image: nvidia/k8s-device-plugin:1.0.0-beta4
-        name: nvidia-device-plugin-ctr
+      - image: cargo.dev.caicloud.xyz/release/nvidia-device-plugin-cus:1.1.0
+        name: nvidia-device-plugin
+        env:
+          - name: NODE_NAME
+            valueFrom:
+              fieldRef:
+                fieldPath: spec.nodeName
         securityContext:
           allowPrivilegeEscalation: false
           capabilities:

nvidia.go

@@ -22,8 +22,7 @@ import (
 	"strings"
 
 	"github.com/NVIDIA/gpu-monitoring-tools/bindings/go/nvml"
-
-	pluginapi "k8s.io/kubernetes/pkg/kubelet/apis/deviceplugin/v1beta1"
+	pluginapi "k8s.io/kubelet/pkg/apis/deviceplugin/v1beta1"
 )
 
 const (
@@ -32,11 +31,11 @@ const (
 )
 
 type ResourceManager interface {
-	Devices() []*pluginapi.Device
+	Devices() ([]*pluginapi.Device, []*nvml.Device)
 	CheckHealth(stop <-chan interface{}, devices []*pluginapi.Device, unhealthy chan<- *pluginapi.Device)
 }
 
-type GpuDeviceManager struct {}
+type GpuDeviceManager struct{}
 
 func check(err error) {
 	if err != nil {
@@ -48,18 +47,20 @@ func NewGpuDeviceManager() *GpuDeviceManager {
 	return &GpuDeviceManager{}
 }
 
-func (g *GpuDeviceManager) Devices() []*pluginapi.Device {
+func (g *GpuDeviceManager) Devices() ([]*pluginapi.Device, []*nvml.Device) {
 	n, err := nvml.GetDeviceCount()
 	check(err)
 
 	var devs []*pluginapi.Device
+	var nvmlDevs []*nvml.Device
 	for i := uint(0); i < n; i++ {
-		d, err := nvml.NewDeviceLite(i)
+		d, err := nvml.NewDevice(i)
 		check(err)
 		devs = append(devs, buildPluginDevice(d))
+		nvmlDevs = append(nvmlDevs, d)
 	}
 
-	return devs
+	return devs, nvmlDevs
 }
 
 func (g *GpuDeviceManager) CheckHealth(stop <-chan interface{}, devices []*pluginapi.Device, unhealthy chan<- *pluginapi.Device) {