host network promoted to spec field

apis/inference/v1alpha1/elasticbatchjob_types.go

@@ -41,6 +41,9 @@ type ElasticBatchJobSpec struct {
 	//      "Worker": ReplicaSpec,
 	//   }
 	ElasticBatchReplicaSpecs map[common.ReplicaType]*common.ReplicaSpec `json:"elasticBatchReplicaSpecs"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *common.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // +genclient

apis/training/v1alpha1/elasticdljob_types.go

@@ -35,6 +35,9 @@ type ElasticDLJobSpec struct {
 	//     "Master": ElasticDLReplicaSpec,
 	//   }
 	ElasticDLReplicaSpecs map[common.ReplicaType]*common.ReplicaSpec `json:"elasticdlReplicaSpecs"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *common.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // +genclient

apis/training/v1alpha1/marsjob_types.go

@@ -40,6 +40,9 @@ type MarsJobSpec struct {
 	// MarsReplicaSpecs is a map of MarsReplicaType(key) to ReplicaSpec(value),
 	// specifying replicas and template of each type.
 	MarsReplicaSpecs map[commonv1.ReplicaType]*commonv1.ReplicaSpec `json:"marsReplicaSpecs"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *commonv1.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // MarsJobStatus defines the observed state of MarsJob

apis/training/v1alpha1/mpijob_types.go

@@ -46,6 +46,9 @@ type MPIJobSpec struct {
 
 	// LegacySpec reserves the deprecated fields for backward compatibility.
 	*MPIJobLegacySpec `json:",inline"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *apiv1.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // MPIJobLegacySpec is a collection of legacy fields that were used in v1alpha1/v1alpha2 but

apis/training/v1alpha1/pytorchjob_types.go

@@ -46,6 +46,9 @@ type PyTorchJobSpec struct {
 	// CacheBackend is used to configure the cache engine for job
 	// +optional
 	CacheBackend *cachev1alpha1.CacheBackendSpec `json:"cacheBackend"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *common.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // PyTorchJobStatus defines the observed state of PyTorchJob

apis/training/v1alpha1/tfjob_types.go

@@ -54,6 +54,9 @@ type TFJobSpec struct {
 	// CacheBackend is used to configure the cache engine for job
 	// +optional
 	CacheBackend *cachev1alpha1.CacheBackendSpec `json:"cacheBackend"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *commonv1.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // +genclient

apis/training/v1alpha1/xdljob_types.go

@@ -48,6 +48,9 @@ type XDLJobSpec struct {
 	// MinFinishWorkPercentage takes precedence over  MinFinishWorkerNum if both are
 	// specified.
 	MinFinishWorkerPercentage *int32 `json:"minFinishWorkRate,omitempty"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *v1.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // XDLJobStatus defines the observed state of XDLJob

apis/training/v1alpha1/xgboostjob_types.go

@@ -37,6 +37,9 @@ type XGBoostJobSpec struct {
 	//     "Worker": ReplicaSpec,
 	//   }
 	XGBReplicaSpecs map[commonv1.ReplicaType]*commonv1.ReplicaSpec `json:"xgbReplicaSpecs"`
+
+	// NetworkMode defines network mode for intra job communicating.
+	NetworkMode *commonv1.NetworkMode `json:"networkmode,omitempty"`
 }
 
 // XGBoostJobStatus defines the observed state of XGBoostJob

config/crd/bases/inference.kubedl.io_elasticbatchjobs.yaml

@@ -3069,6 +3069,8 @@ spec:
                       type: object
                   type: object
                 type: object
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

config/crd/bases/training.kubedl.io_elasticdljobs.yaml

@@ -3063,6 +3063,8 @@ spec:
                       type: object
                   type: object
                 type: object
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

config/crd/bases/training.kubedl.io_marsjobs.yaml

@@ -3063,6 +3063,8 @@ spec:
                       type: object
                   type: object
                 type: object
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

config/crd/bases/training.kubedl.io_mpijobs.yaml

@@ -3075,6 +3075,8 @@ spec:
                       type: object
                   type: object
                 type: object
+              networkmode:
+                type: string
               processingResourceType:
                 type: string
               processingUnits:

config/crd/bases/training.kubedl.io_pytorchjobs.yaml

@@ -161,6 +161,8 @@ spec:
                         type: object
                     type: object
                 type: object
+              networkmode:
+                type: string
               pytorchReplicaSpecs:
                 additionalProperties:
                   properties:

config/crd/bases/training.kubedl.io_tfjobs.yaml

@@ -161,6 +161,8 @@ spec:
                         type: object
                     type: object
                 type: object
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

config/crd/bases/training.kubedl.io_xdljobs.yaml

@@ -72,6 +72,8 @@ spec:
               minFinishWorkRate:
                 format: int32
                 type: integer
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

config/crd/bases/training.kubedl.io_xgboostjobs.yaml

@@ -66,6 +66,8 @@ spec:
                 required:
                 - schedule
                 type: object
+              networkmode:
+                type: string
               schedulingPolicy:
                 properties:
                   minAvailable:

controllers/elasticbatch/elasticbatchjob_controller.go

@@ -130,7 +130,7 @@ func (r *ElasticBatchJobReconciler) Reconcile(_ context.Context, req ctrl.Reques
 	// Set default properties for elasticbatch job.
 	r.scheme.Default(elasticbatchJob)
 
-	result, err := r.ctrl.ReconcileJobs(elasticbatchJob, elasticbatchJob.Spec.ElasticBatchReplicaSpecs, elasticbatchJob.Status, &elasticbatchJob.Spec.RunPolicy, nil, nil)
+	result, err := r.ctrl.ReconcileJobs(elasticbatchJob, elasticbatchJob.Spec.ElasticBatchReplicaSpecs, elasticbatchJob.Status, &elasticbatchJob.Spec.RunPolicy, nil, nil, elasticbatchJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "elasticbatch job reconcile failed")
 		return result, err

controllers/elasticdl/elasticdljob_controller.go

@@ -112,7 +112,7 @@ func (r *ElasticDLJobReconciler) Reconcile(_ context.Context, req ctrl.Request)
 	// Set default properties for elasicdl job.
 	r.scheme.Default(elasticdlJob)
 
-	result, err := r.ctrl.ReconcileJobs(elasticdlJob, elasticdlJob.Spec.ElasticDLReplicaSpecs, elasticdlJob.Status, &elasticdlJob.Spec.RunPolicy, nil, nil)
+	result, err := r.ctrl.ReconcileJobs(elasticdlJob, elasticdlJob.Spec.ElasticDLReplicaSpecs, elasticdlJob.Status, &elasticdlJob.Spec.RunPolicy, nil, nil, elasticdlJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "elasticdl job reconcile failed")
 		return result, err

controllers/mars/marsjob_controller.go

@@ -118,7 +118,7 @@ func (r *MarsJobReconciler) Reconcile(_ context.Context, req ctrl.Request) (ctrl
 
 	r.scheme.Default(marsJob)
 
-	result, err := r.ctrl.ReconcileJobs(marsJob, marsJob.Spec.MarsReplicaSpecs, marsJob.Status.JobStatus, &marsJob.Spec.RunPolicy, nil, nil)
+	result, err := r.ctrl.ReconcileJobs(marsJob, marsJob.Spec.MarsReplicaSpecs, marsJob.Status.JobStatus, &marsJob.Spec.RunPolicy, nil, nil, marsJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "mars job reconcile failed")
 		return result, err

controllers/mpi/mpijob_controller.go

@@ -152,7 +152,7 @@ func (r *MPIJobReconciler) Reconcile(_ context.Context, req ctrl.Request) (ctrl.
 	// Set default properties for tensorflow job.
 	r.scheme.Default(mpiJob)
 
-	result, err = r.ctrl.ReconcileJobs(mpiJob, mpiJob.Spec.MPIReplicaSpecs, mpiJob.Status, &mpiJob.Spec.RunPolicy, nil, nil)
+	result, err = r.ctrl.ReconcileJobs(mpiJob, mpiJob.Spec.MPIReplicaSpecs, mpiJob.Status, &mpiJob.Spec.RunPolicy, nil, nil, mpiJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "mpi job reconcile failed")
 		return result, err

controllers/pytorch/pytorchjob_controller.go

@@ -145,7 +145,7 @@ func (r *PytorchJobReconciler) Reconcile(_ context.Context, req ctrl.Request) (c
 	// Set default properties for pytorch job.
 	r.scheme.Default(pytorchJob)
 
-	result, err := r.ctrl.ReconcileJobs(pytorchJob, pytorchJob.Spec.PyTorchReplicaSpecs, pytorchJob.Status, &pytorchJob.Spec.RunPolicy, pytorchJob.Spec.ModelVersion, pytorchJob.Spec.CacheBackend)
+	result, err := r.ctrl.ReconcileJobs(pytorchJob, pytorchJob.Spec.PyTorchReplicaSpecs, pytorchJob.Status, &pytorchJob.Spec.RunPolicy, pytorchJob.Spec.ModelVersion, pytorchJob.Spec.CacheBackend, pytorchJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "pytorch job reconcile failed")
 		return result, err
@@ -241,7 +241,7 @@ func (r *PytorchJobReconciler) SetClusterSpec(ctx context.Context, job interface
 	}
 
 	masterRole := rtype == strings.ToLower(string(training.PyTorchReplicaTypeMaster))
-	if masterHostPort, ok := job_controller.GetHostNetworkPortFromContext(ctx, "master", "0"); job_controller.EnableHostNetwork(pytorchJob) && ok {
+	if masterHostPort, ok := job_controller.GetHostNetworkPortFromContext(ctx, "master", "0"); job_controller.EnableHostNetwork(pytorchJob.Spec.NetworkMode) && ok {
 		if masterRole || features.KubeDLFeatureGates.Enabled(features.HostNetWithHeadlessSvc) {
 			masterPort = masterHostPort
 		}

controllers/tensorflow/tensorflow.go

@@ -135,7 +135,7 @@ func genClusterSpec(ctx context.Context, tfJob *training.TFJob, selfType, selfIn
 			selfPort := port
 			// Set endpoint port as selected hostnetwork port so that tensorflow worker process could listen
 			// to correct port by TF_CONFIG[cluster].
-			if job_controller.EnableHostNetwork(tfJob) && rt == selfType && strconv.Itoa(int(i)) == selfIndex {
+			if job_controller.EnableHostNetwork(tfJob.Spec.NetworkMode) && rt == selfType && strconv.Itoa(int(i)) == selfIndex {
 				hostPort, ok := job_controller.GetHostNetworkPortFromContext(ctx, selfType, selfIndex)
 				if ok {
 					selfPort = hostPort

controllers/tensorflow/tfjob_controller.go

@@ -171,7 +171,7 @@ func (r *TFJobReconciler) Reconcile(_ context.Context, req ctrl.Request) (ctrl.R
 	// Set default properties for tensorflow job.
 	r.scheme.Default(tfJob)
 
-	result, err := r.ctrl.ReconcileJobs(tfJob, tfJob.Spec.TFReplicaSpecs, tfJob.Status, &tfJob.Spec.RunPolicy, tfJob.Spec.ModelVersion, tfJob.Spec.CacheBackend)
+	result, err := r.ctrl.ReconcileJobs(tfJob, tfJob.Spec.TFReplicaSpecs, tfJob.Status, &tfJob.Spec.RunPolicy, tfJob.Spec.ModelVersion, tfJob.Spec.CacheBackend, tfJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "tensorflow job reconcile failed")
 		return result, err

controllers/xdl/xdljob_controller.go

@@ -131,7 +131,7 @@ func (r *XDLJobReconciler) Reconcile(_ context.Context, request reconcile.Reques
 	// Set default properties for xdl job.
 	r.scheme.Default(xdlJob)
 
-	result, err := r.ctrl.ReconcileJobs(xdlJob, xdlJob.Spec.XDLReplicaSpecs, xdlJob.Status, &xdlJob.Spec.RunPolicy, nil, nil)
+	result, err := r.ctrl.ReconcileJobs(xdlJob, xdlJob.Spec.XDLReplicaSpecs, xdlJob.Status, &xdlJob.Spec.RunPolicy, nil, nil, xdlJob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "xdl job reconcile failed.")
 		return result, err

controllers/xgboost/xgboostjob_controller.go

@@ -118,7 +118,7 @@ func (r *XgboostJobReconciler) Reconcile(_ context.Context, req reconcile.Reques
 	// Set default properties for xgboost job
 	r.scheme.Default(xgboostjob)
 
-	result, err := r.ctrl.ReconcileJobs(xgboostjob, xgboostjob.Spec.XGBReplicaSpecs, xgboostjob.Status.JobStatus, &xgboostjob.Spec.RunPolicy, nil, nil)
+	result, err := r.ctrl.ReconcileJobs(xgboostjob, xgboostjob.Spec.XGBReplicaSpecs, xgboostjob.Status.JobStatus, &xgboostjob.Spec.RunPolicy, nil, nil, xgboostjob.Spec.NetworkMode)
 	if err != nil {
 		log.Error(err, "xgboost job reconcile failed")
 		return result, err

pkg/job_controller/api/v1/constants.go

@@ -30,8 +30,6 @@ const (
 	AnnotationGitSyncConfig = KubeDLPrefix + "/git-sync-config"
 	// AnnotationTenancyInfo annotate tenancy information.
 	AnnotationTenancyInfo = KubeDLPrefix + "/tenancy"
-	// AnnotationNetworkMode annotate job network mode.
-	AnnotationNetworkMode = KubeDLPrefix + "/network-mode"
 	// AnnotationEnableElasticTraining indicates job enables elastic training.
 	AnnotationEnableElasticTraining = KubeDLPrefix + "/enable-elastic-training"
 	// AnnotationElasticScaleState indicates current progress of elastic scaling (inflight | done)
@@ -75,14 +73,6 @@ const (
 	JobReplicaTypeAIMaster ReplicaType = "AIMaster"
 )
 
-// NetworkMode defines network mode for intra job communicating.
-type NetworkMode string
-
-const (
-	// HostNetworkMode indicates that replicas use host-network to communicate with each other.
-	HostNetworkMode NetworkMode = "host"
-)
-
 const (
 	ElasticScaleInflight = "inflight"
 	ElasticScaleDone     = "done"

pkg/job_controller/api/v1/interface.go

@@ -9,6 +9,7 @@ import (
 )
 
 // ControllerInterface defines the Interface to be implemented by custom operators. e.g. tf-operator needs to implement this interface
+// +k8s:deepcopy-gen=false
 type ControllerInterface interface {
 	//ControllerName Returns the Controller name
 	ControllerName() string
@@ -72,6 +73,7 @@ type ControllerInterface interface {
 }
 
 // ElasticScaling defines the interface to be implemented by custom workload elastic behaviors.
+// +k8s:deepcopy-gen=false
 type ElasticScaling interface {
 	// EnableElasticScaling indicates workload enables elastic scaling or not.
 	EnableElasticScaling(job v1.Object, runPolicy *RunPolicy) bool

pkg/job_controller/api/v1/types.go

@@ -312,3 +312,11 @@ type DAGCondition struct {
 	// OnPhase defines at which phase the upstream replica will trigger this condition.
 	OnPhase v1.PodPhase `json:"onPhase"`
 }
+
+// NetworkMode defines network mode for intra job communicating.
+type NetworkMode string
+
+const (
+	// HostNetworkMode indicates that replicas use host-network to communicate with each other.
+	HostNetworkMode NetworkMode = "host"
+)