Merge pull request kubernetes#24836 from Clarifai/gpu-impl

Automatic merge from submit-queue

WIP v0 NVIDIA GPU support

```release-note
* Alpha support for scheduling pods on machines with NVIDIA GPUs whose kubelets use the `--experimental-nvidia-gpus` flag, using the alpha.kubernetes.io/nvidia-gpu resource 
```

Implements part of kubernetes#24071 for  kubernetes#23587

I am not familiar with the scheduler enough to know what to do with the scores. Mostly punting for now.

Missing items from the implementation plan: limitranger, rkt support, kubectl
support and docs

cc @erictune @davidopp @dchen1107 @vishh @Hui-Zhi @gopinatht

cmd/kubelet/app/options/options.go

@@ -106,6 +106,7 @@ func NewKubeletServer() *KubeletServer {
 			MaxPerPodContainerCount:     2,
 			MaxOpenFiles:                1000000,
 			MaxPods:                     110,
+			NvidiaGPUs:                  0,
 			MinimumGCAge:                unversioned.Duration{Duration: 1 * time.Minute},
 			NetworkPluginDir:            "/usr/libexec/kubernetes/kubelet-plugins/net/exec/",
 			NetworkPluginName:           "",
@@ -227,6 +228,7 @@ func (s *KubeletServer) AddFlags(fs *pflag.FlagSet) {
 	fs.BoolVar(&s.BabysitDaemons, "babysit-daemons", s.BabysitDaemons, "If true, the node has babysitter process monitoring docker and kubelet.")
 	fs.MarkDeprecated("babysit-daemons", "Will be removed in a future version.")
 	fs.Int32Var(&s.MaxPods, "max-pods", s.MaxPods, "Number of Pods that can run on this Kubelet.")
+	fs.Int32Var(&s.NvidiaGPUs, "experimental-nvidia-gpus", s.NvidiaGPUs, "Number of NVIDIA GPU devices on this node. Only 0 (default) and 1 are currently supported.")
 	fs.StringVar(&s.DockerExecHandlerName, "docker-exec-handler", s.DockerExecHandlerName, "Handler to use when executing a command in a container. Valid values are 'native' and 'nsenter'. Defaults to 'native'.")
 	fs.StringVar(&s.NonMasqueradeCIDR, "non-masquerade-cidr", s.NonMasqueradeCIDR, "Traffic to IPs outside this range will use IP masquerade.")
 	fs.StringVar(&s.PodCIDR, "pod-cidr", "", "The CIDR to use for pod IP addresses, only used in standalone mode.  In cluster mode, this is obtained from the master.")

cmd/kubelet/app/server.go

@@ -228,6 +228,7 @@ func UnsecuredKubeletConfig(s *options.KubeletServer) (*KubeletConfig, error) {
 		MaxOpenFiles:              s.MaxOpenFiles,
 		MaxPerPodContainerCount:   int(s.MaxPerPodContainerCount),
 		MaxPods:                   int(s.MaxPods),
+		NvidiaGPUs:                int(s.NvidiaGPUs),
 		MinimumGCAge:              s.MinimumGCAge.Duration,
 		Mounter:                   mounter,
 		NetworkPluginName:         s.NetworkPluginName,
@@ -558,6 +559,7 @@ func SimpleKubelet(client *clientset.Clientset,
 		MaxOpenFiles:              1024,
 		MaxPerPodContainerCount:   2,
 		MaxPods:                   maxPods,
+		NvidiaGPUs:                0,
 		MinimumGCAge:              minimumGCAge,
 		Mounter:                   mount.New(),
 		NodeStatusUpdateFrequency: nodeStatusUpdateFrequency,
@@ -750,6 +752,7 @@ type KubeletConfig struct {
 	NodeLabels                     map[string]string
 	NodeStatusUpdateFrequency      time.Duration
 	NonMasqueradeCIDR              string
+	NvidiaGPUs                     int
 	OOMAdjuster                    *oom.OOMAdjuster
 	OSInterface                    kubecontainer.OSInterface
 	PodCIDR                        string
@@ -860,6 +863,7 @@ func CreateAndInitKubelet(kc *KubeletConfig) (k KubeletBootstrap, pc *config.Pod
 		kc.PodCIDR,
 		kc.ReconcileCIDR,
 		kc.MaxPods,
+		kc.NvidiaGPUs,
 		kc.DockerExecHandler,
 		kc.ResolverConfig,
 		kc.CPUCFSQuota,

docs/admin/kubelet.md

@@ -91,6 +91,7 @@ kubelet
       --eviction-soft="": A set of eviction thresholds (e.g. memory.available<1.5Gi) that if met over a corresponding grace period would trigger a pod eviction.
       --eviction-soft-grace-period="": A set of eviction grace periods (e.g. memory.available=1m30s) that correspond to how long a soft eviction threshold must hold before triggering a pod eviction.
       --experimental-flannel-overlay[=false]: Experimental support for starting the kubelet with the default overlay network (flannel). Assumes flanneld is already running in client mode. [default=false]
+      --experimental-nvidia-gpus=0: Number of NVIDIA GPU devices on this node. Only 0 (default) and 1 are currently supported.
       --file-check-frequency=20s: Duration between checking config files for new data
       --google-json-key="": The Google Cloud Platform Service Account JSON Key to use for authentication.
       --hairpin-mode="promiscuous-bridge": How should the kubelet setup hairpin NAT. This allows endpoints of a Service to loadbalance back to themselves if they should try to access their own Service. Valid values are "promiscuous-bridge", "hairpin-veth" and "none".

hack/verify-flags/known-flags.txt

@@ -127,6 +127,7 @@ executor-path
 executor-suicide-timeout
 experimental-flannel-overlay
 experimental-keystone-url
+experimental-nvidia-gpus
 experimental-prefix
 external-hostname
 external-ip

pkg/api/resource_helpers.go

@@ -49,6 +49,13 @@ func (self *ResourceList) Pods() *resource.Quantity {
 	return &resource.Quantity{}
 }
 
+func (self *ResourceList) NvidiaGPU() *resource.Quantity {
+	if val, ok := (*self)[ResourceNvidiaGPU]; ok {
+		return &val
+	}
+	return &resource.Quantity{}
+}
+
 func GetContainerStatus(statuses []ContainerStatus, name string) (ContainerStatus, bool) {
 	for i := range statuses {
 		if statuses[i].Name == name {

pkg/api/types.go

@@ -1922,13 +1922,20 @@ type NodeResources struct {
 // ResourceName is the name identifying various resources in a ResourceList.
 type ResourceName string
 
+// Resource names must be not more than 63 characters, consisting of upper- or lower-case alphanumeric characters,
+// with the -, _, and . characters allowed anywhere, except the first or last character.
+// The default convention, matching that for annotations, is to use lower-case names, with dashes, rather than
+// camel case, separating compound words.
+// Fully-qualified resource typenames are constructed from a DNS-style subdomain, followed by a slash `/` and a name.
 const (
 	// CPU, in cores. (500m = .5 cores)
 	ResourceCPU ResourceName = "cpu"
 	// Memory, in bytes. (500Gi = 500GiB = 500 * 1024 * 1024 * 1024)
 	ResourceMemory ResourceName = "memory"
 	// Volume size, in bytes (e,g. 5Gi = 5GiB = 5 * 1024 * 1024 * 1024)
 	ResourceStorage ResourceName = "storage"
+	// NVIDIA GPU, in devices. Alpha, might change: although fractional and allowing values >1, only one whole device per node is assigned.
+	ResourceNvidiaGPU ResourceName = "alpha.kubernetes.io/nvidia-gpu"
 	// Number of Pods that may be running on this Node: see ResourcePods
 )
 

pkg/api/v1/types.go

@@ -2307,13 +2307,21 @@ type NodeAddress struct {
 // ResourceName is the name identifying various resources in a ResourceList.
 type ResourceName string
 
+// Resource names must be not more than 63 characters, consisting of upper- or lower-case alphanumeric characters,
+// with the -, _, and . characters allowed anywhere, except the first or last character.
+// The default convention, matching that for annotations, is to use lower-case names, with dashes, rather than
+// camel case, separating compound words.
+// Fully-qualified resource typenames are constructed from a DNS-style subdomain, followed by a slash `/` and a name.
 const (
 	// CPU, in cores. (500m = .5 cores)
 	ResourceCPU ResourceName = "cpu"
 	// Memory, in bytes. (500Gi = 500GiB = 500 * 1024 * 1024 * 1024)
 	ResourceMemory ResourceName = "memory"
 	// Volume size, in bytes (e,g. 5Gi = 5GiB = 5 * 1024 * 1024 * 1024)
 	ResourceStorage ResourceName = "storage"
+	// NVIDIA GPU, in devices. Alpha, might change: although fractional and allowing values >1, only one whole device per node is assigned.
+	ResourceNvidiaGPU ResourceName = "alpha.kubernetes.io/nvidia-gpu"
+	// Number of Pods that may be running on this Node: see ResourcePods
 )
 
 // ResourceList is a set of (resource name, quantity) pairs.

pkg/api/validation/validation.go

@@ -2489,6 +2489,7 @@ func validateBasicResource(quantity resource.Quantity, fldPath *field.Path) fiel
 func ValidateResourceRequirements(requirements *api.ResourceRequirements, fldPath *field.Path) field.ErrorList {
 	allErrs := field.ErrorList{}
 	limPath := fldPath.Child("limits")
+	reqPath := fldPath.Child("requests")
 	for resourceName, quantity := range requirements.Limits {
 		fldPath := limPath.Key(string(resourceName))
 		// Validate resource name.
@@ -2499,12 +2500,14 @@ func ValidateResourceRequirements(requirements *api.ResourceRequirements, fldPat
 		// Check that request <= limit.
 		requestQuantity, exists := requirements.Requests[resourceName]
 		if exists {
-			if quantity.Cmp(requestQuantity) < 0 {
+			// For GPUs, require that no request be set.
+			if resourceName == api.ResourceNvidiaGPU {
+				allErrs = append(allErrs, field.Invalid(reqPath, requestQuantity.String(), "cannot be set"))
+			} else if quantity.Cmp(requestQuantity) < 0 {
 				allErrs = append(allErrs, field.Invalid(fldPath, quantity.String(), "must be greater than or equal to request"))
 			}
 		}
 	}
-	reqPath := fldPath.Child("requests")
 	for resourceName, quantity := range requirements.Requests {
 		fldPath := reqPath.Key(string(resourceName))
 		// Validate resource name.

pkg/api/validation/validation_test.go

@@ -1383,6 +1383,22 @@ func TestValidateContainers(t *testing.T) {
 			},
 			ImagePullPolicy: "IfNotPresent",
 		},
+		{
+			Name:  "resources-test-with-gpu",
+			Image: "image",
+			Resources: api.ResourceRequirements{
+				Requests: api.ResourceList{
+					api.ResourceName(api.ResourceCPU):    resource.MustParse("10"),
+					api.ResourceName(api.ResourceMemory): resource.MustParse("10G"),
+				},
+				Limits: api.ResourceList{
+					api.ResourceName(api.ResourceCPU):       resource.MustParse("10"),
+					api.ResourceName(api.ResourceMemory):    resource.MustParse("10G"),
+					api.ResourceName(api.ResourceNvidiaGPU): resource.MustParse("1"),
+				},
+			},
+			ImagePullPolicy: "IfNotPresent",
+		},
 		{
 			Name:  "resources-request-limit-simple",
 			Image: "image",
@@ -1606,6 +1622,25 @@ func TestValidateContainers(t *testing.T) {
 				ImagePullPolicy: "IfNotPresent",
 			},
 		},
+		"Resource can only have GPU limit": {
+			{
+				Name:  "resources-request-limit-edge",
+				Image: "image",
+				Resources: api.ResourceRequirements{
+					Requests: api.ResourceList{
+						api.ResourceName(api.ResourceCPU):       resource.MustParse("10"),
+						api.ResourceName(api.ResourceMemory):    resource.MustParse("10G"),
+						api.ResourceName(api.ResourceNvidiaGPU): resource.MustParse("1"),
+					},
+					Limits: api.ResourceList{
+						api.ResourceName(api.ResourceCPU):       resource.MustParse("10"),
+						api.ResourceName(api.ResourceMemory):    resource.MustParse("10G"),
+						api.ResourceName(api.ResourceNvidiaGPU): resource.MustParse("1"),
+					},
+				},
+				ImagePullPolicy: "IfNotPresent",
+			},
+		},
 		"Request limit simple invalid": {
 			{
 				Name:  "abc-123",

pkg/apis/componentconfig/deep_copy_generated.go

@@ -276,6 +276,7 @@ func DeepCopy_componentconfig_KubeletConfiguration(in KubeletConfiguration, out
 	out.HairpinMode = in.HairpinMode
 	out.BabysitDaemons = in.BabysitDaemons
 	out.MaxPods = in.MaxPods
+	out.NvidiaGPUs = in.NvidiaGPUs
 	out.DockerExecHandlerName = in.DockerExecHandlerName
 	out.PodCIDR = in.PodCIDR
 	out.ResolverConfig = in.ResolverConfig