use the priority of kube-batch

manifests/crd.yaml

@@ -23,6 +23,8 @@ spec:
       properties:
         spec:
           properties:
+            priorityClassName:
+              type: string
             pytorchReplicaSpecs:
               properties:
                 Master:

manifests/podgroup.yaml

@@ -22,6 +22,8 @@ spec:
             minMember:
               format: int32
               type: integer
+            priorityClassName:
+              type: string
           type: object
         status:
           properties:

pkg/apis/pytorch/v1/types.go

@@ -69,6 +69,10 @@ type PyTorchJobSpec struct {
 	//     "Worker": PyTorchReplicaSpec,
 	//   }
 	PyTorchReplicaSpecs map[PyTorchReplicaType]*common.ReplicaSpec `json:"pytorchReplicaSpecs"`
+
+	//添加判断优先级的属性
+	//add PriorityClassName
+	PriorityClassName string `json:"priorityClassName,omitempty"`
 }
 
 // PyTorchReplicaType is the type for PyTorchReplica. Can be one of "Master" or "Worker".

pkg/controller.v1/pytorch/controller.go

@@ -437,7 +437,9 @@ func (pc *PyTorchController) reconcilePyTorchJobs(job *pyv1.PyTorchJob) error {
 
 	if pc.Config.EnableGangScheduling {
 		minAvailableReplicas := getTotalReplicas(job)
-		_, err := pc.SyncPodGroup(job, minAvailableReplicas)
+		priorityClassName:=getPriorityClassName(job)
+		//_, err := pc.SyncPodGroup(job, minAvailableReplicas)
+		_, err := pc.SyncPodGroupTest(job, minAvailableReplicas,priorityClassName)
 		if err != nil {
 			logger.Warnf("Sync PodGroup %v: %v", job.Name, err)
 		}

pkg/controller.v1/pytorch/job.go

@@ -216,3 +216,10 @@ func getTotalFailedReplicas(job *pyv1.PyTorchJob) int32 {
 	}
 	return totalFailedReplicas
 }
+
+func getPriorityClassName(job *pyv1.PyTorchJob) string {
+	var priorityClassName string
+	priorityClassName=job.Spec.PriorityClassName
+
+	return priorityClassName
+}

pkg/controller.v1/pytorch/status.go

@@ -69,6 +69,7 @@ func (pc *PyTorchController) updateStatusSingle(job *pyv1.PyTorchJob, rtype pyv1
 
 	// Expect to have `replicas - succeeded` pods alive.
 	commonType := common.ReplicaType(rtype)
+	//expected是成功的判断标志，等于0时，成功的数量等于副本数，认为成功
 	expected := replicas - int(job.Status.ReplicaStatuses[commonType].Succeeded)
 	running := int(job.Status.ReplicaStatuses[commonType].Active)
 	failed := int(job.Status.ReplicaStatuses[commonType].Failed)