memory manager: add memory manager flag under kubelet options and kubelet config

The commit also includes generated files after `make generated_files`.

Signed-off-by: Byonggon Chun <bg.chun@samsung.com>

api/api-rules/violation_exceptions.list

@@ -392,6 +392,7 @@ API rule violation: list_type_missing,k8s.io/kubelet/config/v1alpha1,CredentialP
 API rule violation: list_type_missing,k8s.io/kubelet/config/v1beta1,KubeletConfiguration,AllowedUnsafeSysctls
 API rule violation: list_type_missing,k8s.io/kubelet/config/v1beta1,KubeletConfiguration,ClusterDNS
 API rule violation: list_type_missing,k8s.io/kubelet/config/v1beta1,KubeletConfiguration,EnforceNodeAllocatable
+API rule violation: list_type_missing,k8s.io/kubelet/config/v1beta1,KubeletConfiguration,ReservedMemory
 API rule violation: list_type_missing,k8s.io/kubelet/config/v1beta1,KubeletConfiguration,TLSCipherSuites
 API rule violation: list_type_missing,k8s.io/metrics/pkg/apis/metrics/v1alpha1,PodMetrics,Containers
 API rule violation: list_type_missing,k8s.io/metrics/pkg/apis/metrics/v1beta1,PodMetrics,Containers

cmd/kubelet/app/options/options.go

@@ -550,4 +550,9 @@ Runtime log sanitization may introduce significant computation overhead and ther
 
 	// Graduated experimental flags, kept for backward compatibility
 	fs.BoolVar(&c.KernelMemcgNotification, "experimental-kernel-memcg-notification", c.KernelMemcgNotification, "Use kernelMemcgNotification configuration, this flag will be removed in 1.23.")
+
+	// Memory Manager Flags
+	fs.StringVar(&c.MemoryManagerPolicy, "memory-manager-policy", c.MemoryManagerPolicy, "Memory Manager policy to use. Possible values: 'none', 'static'. Default: 'none'")
+	// TODO: once documentation link is available, replace KEP link with the documentation one.
+	fs.Var(cliflag.NewBracketSeparatedSliceMapStringString(&c.ReservedMemory), "reserved-memory", "A comma separated list of bracket-enclosed configuration for memory manager (e.g. {numa-node=0, type=memory, limit=1Gi}, {numa-node=1, type=memory, limit=1Gi}). The total sum for each memory type should be equal to the sum of kube-reserved, system-reserved and eviction-threshold. See more details under https://github.com/kubernetes/enhancements/tree/master/keps/sig-node/1769-memory-manager#reserved-memory-flag")
 }

cmd/kubelet/app/server.go

@@ -71,6 +71,7 @@ import (
 	"k8s.io/kubernetes/cmd/kubelet/app/options"
 	"k8s.io/kubernetes/pkg/api/legacyscheme"
 	api "k8s.io/kubernetes/pkg/apis/core"
+	corev1helper "k8s.io/kubernetes/pkg/apis/core/v1/helper"
 	"k8s.io/kubernetes/pkg/capabilities"
 	"k8s.io/kubernetes/pkg/credentialprovider"
 	"k8s.io/kubernetes/pkg/features"
@@ -687,6 +688,12 @@ func run(ctx context.Context, s *options.KubeletServer, kubeDeps *kubelet.Depend
 			s.SystemReserved["cpu"] = strconv.Itoa(reservedSystemCPUs.Size())
 			klog.Infof("After cpu setting is overwritten, KubeReserved=\"%v\", SystemReserved=\"%v\"", s.KubeReserved, s.SystemReserved)
 		}
+
+		reservedMemory, err := parseReservedMemoryConfig(s.ReservedMemory)
+		if err != nil {
+			return err
+		}
+
 		kubeReserved, err := parseResourceList(s.KubeReserved)
 		if err != nil {
 			return err
@@ -732,14 +739,16 @@ func run(ctx context.Context, s *options.KubeletServer, kubeDeps *kubelet.Depend
 					ReservedSystemCPUs:       reservedSystemCPUs,
 					HardEvictionThresholds:   hardEvictionThresholds,
 				},
-				QOSReserved:                           *experimentalQOSReserved,
-				ExperimentalCPUManagerPolicy:          s.CPUManagerPolicy,
-				ExperimentalCPUManagerReconcilePeriod: s.CPUManagerReconcilePeriod.Duration,
-				ExperimentalPodPidsLimit:              s.PodPidsLimit,
-				EnforceCPULimits:                      s.CPUCFSQuota,
-				CPUCFSQuotaPeriod:                     s.CPUCFSQuotaPeriod.Duration,
-				ExperimentalTopologyManagerPolicy:     s.TopologyManagerPolicy,
-				ExperimentalTopologyManagerScope:      s.TopologyManagerScope,
+				QOSReserved:                             *experimentalQOSReserved,
+				ExperimentalCPUManagerPolicy:            s.CPUManagerPolicy,
+				ExperimentalCPUManagerReconcilePeriod:   s.CPUManagerReconcilePeriod.Duration,
+				ExperimentalMemoryManagerPolicy:         s.MemoryManagerPolicy,
+				ExperimentalMemoryManagerReservedMemory: reservedMemory,
+				ExperimentalPodPidsLimit:                s.PodPidsLimit,
+				EnforceCPULimits:                        s.CPUCFSQuota,
+				CPUCFSQuotaPeriod:                       s.CPUCFSQuotaPeriod.Duration,
+				ExperimentalTopologyManagerPolicy:       s.TopologyManagerPolicy,
+				ExperimentalTopologyManagerScope:        s.TopologyManagerScope,
 			},
 			s.FailSwapOn,
 			devicePluginEnabled,
@@ -1296,6 +1305,59 @@ func parseResourceList(m map[string]string) (v1.ResourceList, error) {
 	return rl, nil
 }
 
+func parseReservedMemoryConfig(config []map[string]string) (map[int]map[v1.ResourceName]resource.Quantity, error) {
+	if len(config) == 0 {
+		return nil, nil
+	}
+
+	const (
+		indexKey = "numa-node"
+		typeKey  = "type"
+		limitKey = "limit"
+	)
+
+	keys := []string{indexKey, typeKey, limitKey}
+
+	// check whether all keys are present
+	for _, m := range config {
+		for _, key := range keys {
+			if _, exist := m[key]; !exist {
+				return nil, fmt.Errorf("key: %s is missing in given ReservedMemory flag: %v", key, config)
+			}
+		}
+	}
+
+	parsed := make(map[int]map[v1.ResourceName]resource.Quantity, len(config))
+	for _, m := range config {
+		idxInString, _ := m[indexKey]
+		idx, err := strconv.Atoi(idxInString)
+		if err != nil || idx < 0 {
+			return nil, fmt.Errorf("NUMA index conversion error for value: \"%s\"", idxInString)
+		}
+
+		typeInString, _ := m[typeKey]
+		v1Type := v1.ResourceName(typeInString)
+		if v1Type != v1.ResourceMemory && !corev1helper.IsHugePageResourceName(v1Type) {
+			return nil, fmt.Errorf("memory type conversion error, unknown type: \"%s\"", typeInString)
+		}
+		if corev1helper.IsHugePageResourceName(v1Type) {
+			if _, err := corev1helper.HugePageSizeFromResourceName(v1Type); err != nil {
+				return nil, fmt.Errorf("memory type conversion error, unknown type: \"%s\"", typeInString)
+			}
+		}
+
+		limitInString, _ := m[limitKey]
+		limit, err := resource.ParseQuantity(limitInString)
+		if err != nil || limit.Sign() != 1 {
+			return nil, fmt.Errorf("memory limit conversion error for value \"%s\"", limitInString)
+		}
+		parsed[idx] = make(map[v1.ResourceName]resource.Quantity)
+		parsed[idx][v1Type] = limit
+	}
+
+	return parsed, nil
+}
+
 // BootstrapKubeletConfigController constructs and bootstrap a configuration controller
 func BootstrapKubeletConfigController(dynamicConfigDir string, transform dynamickubeletconfig.TransformFunc) (*kubeletconfiginternal.KubeletConfiguration, *dynamickubeletconfig.Controller, error) {
 	if !utilfeature.DefaultFeatureGate.Enabled(features.DynamicKubeletConfig) {

pkg/features/kube_features.go

@@ -123,6 +123,12 @@ const (
 	// Enable resource managers to make NUMA aligned decisions
 	TopologyManager featuregate.Feature = "TopologyManager"
 
+	// owner: @cynepco3hahue(alukiano) @cezaryzukowski @k-wiatrzyk
+	// alpha:: v1.20
+
+	// Allows to set containers memory affinity according to NUMA topology
+	MemoryManager featuregate.Feature = "MemoryManager"
+
 	// owner: @sjenning
 	// beta: v1.11
 	//

pkg/kubelet/apis/config/types.go

@@ -224,6 +224,9 @@ type KubeletConfiguration struct {
 	// CPU Manager reconciliation period.
 	// Requires the CPUManager feature gate to be enabled.
 	CPUManagerReconcilePeriod metav1.Duration
+	// MemoryManagerPolicy is the name of the policy to use.
+	// Requires the MemoryManager feature gate to be enabled.
+	MemoryManagerPolicy string
 	// TopologyManagerPolicy is the name of the policy to use.
 	// Policies other than "none" require the TopologyManager feature gate to be enabled.
 	TopologyManagerPolicy string
@@ -382,6 +385,12 @@ type KubeletConfiguration struct {
 	// Defaults to 10 seconds, requires GracefulNodeShutdown feature gate to be enabled.
 	// For example, if ShutdownGracePeriod=30s, and ShutdownGracePeriodCriticalPods=10s, during a node shutdown the first 20 seconds would be reserved for gracefully terminating normal pods, and the last 10 seconds would be reserved for terminating critical pods.
 	ShutdownGracePeriodCriticalPods metav1.Duration
+	// A comma separated list of bracket-enclosed configurations for memory manager.
+	// Each configuration describes pre-reserved memory for the particular memory type on a specific NUMA node.
+	// The Memory Manager validates whether total amount of pre-reserved memory is identical to reserved-memory by the Node Allocatable feature.
+	// The format is {numa-node=integer, memory-type=string, limit=string}
+	// (e.g. {numa-node=0, type=memory, limit=1Gi}, {numa-node=1, type=memory, limit=1Gi})
+	ReservedMemory []map[string]string
 }
 
 // KubeletAuthorizationMode denotes the authorization mode for the kubelet

pkg/kubelet/apis/config/v1beta1/zz_generated.conversion.go

@@ -274,6 +274,7 @@ func autoConvert_v1beta1_KubeletConfiguration_To_config_KubeletConfiguration(in
 	out.CgroupDriver = in.CgroupDriver
 	out.CPUManagerPolicy = in.CPUManagerPolicy
 	out.CPUManagerReconcilePeriod = in.CPUManagerReconcilePeriod
+	out.MemoryManagerPolicy = in.MemoryManagerPolicy
 	out.TopologyManagerPolicy = in.TopologyManagerPolicy
 	out.TopologyManagerScope = in.TopologyManagerScope
 	out.QOSReserved = *(*map[string]string)(unsafe.Pointer(&in.QOSReserved))
@@ -352,6 +353,7 @@ func autoConvert_v1beta1_KubeletConfiguration_To_config_KubeletConfiguration(in
 	}
 	out.ShutdownGracePeriod = in.ShutdownGracePeriod
 	out.ShutdownGracePeriodCriticalPods = in.ShutdownGracePeriodCriticalPods
+	out.ReservedMemory = *(*[]map[string]string)(unsafe.Pointer(&in.ReservedMemory))
 	return nil
 }
 
@@ -429,6 +431,7 @@ func autoConvert_config_KubeletConfiguration_To_v1beta1_KubeletConfiguration(in
 	out.CgroupDriver = in.CgroupDriver
 	out.CPUManagerPolicy = in.CPUManagerPolicy
 	out.CPUManagerReconcilePeriod = in.CPUManagerReconcilePeriod
+	out.MemoryManagerPolicy = in.MemoryManagerPolicy
 	out.TopologyManagerPolicy = in.TopologyManagerPolicy
 	out.TopologyManagerScope = in.TopologyManagerScope
 	out.QOSReserved = *(*map[string]string)(unsafe.Pointer(&in.QOSReserved))
@@ -505,6 +508,7 @@ func autoConvert_config_KubeletConfiguration_To_v1beta1_KubeletConfiguration(in
 	}
 	out.ShutdownGracePeriod = in.ShutdownGracePeriod
 	out.ShutdownGracePeriodCriticalPods = in.ShutdownGracePeriodCriticalPods
+	out.ReservedMemory = *(*[]map[string]string)(unsafe.Pointer(&in.ReservedMemory))
 	return nil
 }
 

pkg/kubelet/apis/config/zz_generated.deepcopy.go

@@ -273,6 +273,19 @@ func (in *KubeletConfiguration) DeepCopyInto(out *KubeletConfiguration) {
 	out.Logging = in.Logging
 	out.ShutdownGracePeriod = in.ShutdownGracePeriod
 	out.ShutdownGracePeriodCriticalPods = in.ShutdownGracePeriodCriticalPods
+	if in.ReservedMemory != nil {
+		in, out := &in.ReservedMemory, &out.ReservedMemory
+		*out = make([]map[string]string, len(*in))
+		for i := range *in {
+			if (*in)[i] != nil {
+				in, out := &(*in)[i], &(*out)[i]
+				*out = make(map[string]string, len(*in))
+				for key, val := range *in {
+					(*out)[key] = val
+				}
+			}
+		}
+	}
 	return
 }
 

staging/src/k8s.io/kubelet/config/v1beta1/types.go

@@ -73,6 +73,13 @@ const (
 	// PodTopologyManagerScope represents that
 	// topology policy is applied on a per-pod basis.
 	PodTopologyManagerScope = "pod"
+	// NoneMemoryManagerPolicy is a memory manager none policy, under the none policy
+	// the memory manager will not pin containers memory of guaranteed pods
+	NoneMemoryManagerPolicy = "none"
+	// StaticMemoryManagerPolicy is a memory manager static policy, under the static policy
+	// the memory manager will try to pin containers memory of guaranteed pods to the smallest
+	// possible sub-set of NUMA nodes
+	StaticMemoryManagerPolicy = "static"
 )
 
 // +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object
@@ -433,6 +440,13 @@ type KubeletConfiguration struct {
 	// Default: "10s"
 	// +optional
 	CPUManagerReconcilePeriod metav1.Duration `json:"cpuManagerReconcilePeriod,omitempty"`
+	// MemoryManagerPolicy is the name of the policy to use by memory manager.
+	// Requires the MemoryManager feature gate to be enabled.
+	// Dynamic Kubelet Config (beta): This field should not be updated without a full node
+	// reboot. It is safest to keep this value the same as the local config.
+	// Default: "none"
+	// +optional
+	MemoryManagerPolicy string `json:"memoryManagerPolicy,omitempty"`
 	// TopologyManagerPolicy is the name of the policy to use.
 	// Policies other than "none" require the TopologyManager feature gate to be enabled.
 	// Dynamic Kubelet Config (beta): This field should not be updated without a full node
@@ -824,6 +838,14 @@ type KubeletConfiguration struct {
 	// Default: "10s"
 	// +optional
 	ShutdownGracePeriodCriticalPods metav1.Duration `json:"shutdownGracePeriodCriticalPods,omitempty"`
+	// A comma separated list of bracket-enclosed configurations for memory manager.
+	// Each configuration describes pre-reserved memory for the certain memory type on a specific NUMA node.
+	// The Memory Manager validates whether total amount of pre-reserved memory is identical to reserved-memory by the Node Allocatable feature.
+	// The format is {numa-node=integer, memory-type=string, limit=string}
+	// (e.g. {numa-node=0, type=memory, limit=1Gi}, {numa-node=1, type=memory, limit=1Gi})
+	// Default: nil
+	// +optional
+	ReservedMemory []map[string]string `json:"reservedMemory,omitempty"`
 }
 
 type KubeletAuthorizationMode string

staging/src/k8s.io/kubelet/config/v1beta1/zz_generated.deepcopy.go

@@ -303,6 +303,19 @@ func (in *KubeletConfiguration) DeepCopyInto(out *KubeletConfiguration) {
 	}
 	out.ShutdownGracePeriod = in.ShutdownGracePeriod
 	out.ShutdownGracePeriodCriticalPods = in.ShutdownGracePeriodCriticalPods
+	if in.ReservedMemory != nil {
+		in, out := &in.ReservedMemory, &out.ReservedMemory
+		*out = make([]map[string]string, len(*in))
+		for i := range *in {
+			if (*in)[i] != nil {
+				in, out := &(*in)[i], &(*out)[i]
+				*out = make(map[string]string, len(*in))
+				for key, val := range *in {
+					(*out)[key] = val
+				}
+			}
+		}
+	}
 	return
 }