diff --git a/app/core.go b/app/core.go
index d02abb05..fe8b5014 100644
--- a/app/core.go
+++ b/app/core.go
@@ -85,8 +85,10 @@ func startServiceController(ctx context.Context, initContext ControllerInitConte
 		completedConfig.ClientBuilder.ClientOrDie(initContext.ClientName),
 		completedConfig.SharedInformers.Core().V1().Services(),
 		completedConfig.SharedInformers.Core().V1().Nodes(),
+		completedConfig.SharedInformers.Core().V1().Endpoints(),
 		completedConfig.ComponentConfig.KubeCloudShared.ClusterName,
 		utilfeature.DefaultFeatureGate,
+		completedConfig.ComponentConfig.ServiceController.DirectPodIP,
 	)
 	if err != nil {
 		// This error shouldn't fail. It lives like this as a legacy.
diff --git a/controllers/service/config/types.go b/controllers/service/config/types.go
index 3afca163..1c6a556a 100644
--- a/controllers/service/config/types.go
+++ b/controllers/service/config/types.go
@@ -22,4 +22,9 @@ type ServiceControllerConfiguration struct {
 	// allowed to sync concurrently. Larger number = more responsive service
 	// management, but more CPU (and network) load.
 	ConcurrentServiceSyncs int32
+
+	// directPodIp is the mode of load balancer service
+	// If true configured, load balancer will use pods ip addresses
+	// as pool members instead of nodes addresses.
+	DirectPodIP bool
 }
diff --git a/controllers/service/controller.go b/controllers/service/controller.go
index ae717e6b..063b1848 100644
--- a/controllers/service/controller.go
+++ b/controllers/service/controller.go
@@ -90,6 +90,11 @@ type Controller struct {
 	// services that need to be synced
 	queue workqueue.RateLimitingInterface
 
+	endpointLister       corelisters.EndpointsLister
+	endpointListerSynced cache.InformerSynced
+	// endpoints that need to be synced
+	endpointsQueue workqueue.RateLimitingInterface
+
 	// nodeSyncLock ensures there is only one instance of triggerNodeSync getting executed at one time
 	// and protects internal states (needFullSync) of nodeSync
 	nodeSyncLock sync.Mutex
@@ -97,6 +102,9 @@ type Controller struct {
 	nodeSyncCh chan interface{}
 	// needFullSync indicates if the nodeSyncInternal will do a full node sync on all LB services.
 	needFullSync bool
+
+	// if service controller will create lb svc in directPodIP mode or nodePort mode
+	directPodIP bool
 }
 
 // New returns a new service controller to keep cloud provider service resources
@@ -106,8 +114,10 @@ func New(
 	kubeClient clientset.Interface,
 	serviceInformer coreinformers.ServiceInformer,
 	nodeInformer coreinformers.NodeInformer,
+	endpointInformer coreinformers.EndpointsInformer,
 	clusterName string,
 	featureGate featuregate.FeatureGate,
+	directPodIP bool,
 ) (*Controller, error) {
 	broadcaster := record.NewBroadcaster()
 	broadcaster.StartStructuredLogging(0)
@@ -134,6 +144,10 @@ func New(
 		queue:            workqueue.NewNamedRateLimitingQueue(workqueue.NewItemExponentialFailureRateLimiter(minRetryDelay, maxRetryDelay), "service"),
 		// nodeSyncCh has a size 1 buffer. Only one pending sync signal would be cached.
 		nodeSyncCh: make(chan interface{}, 1),
+		endpointLister:   endpointInformer.Lister(),
+		endpointListerSynced: endpointInformer.Informer().HasSynced,
+		endpointsQueue:   workqueue.NewNamedRateLimitingQueue(workqueue.NewItemExponentialFailureRateLimiter(minRetryDelay, maxRetryDelay), "endpoint"),
+		directPodIP: directPodIP,
 	}
 
 	serviceInformer.Informer().AddEventHandlerWithResyncPeriod(
@@ -161,6 +175,10 @@ func New(
 	s.serviceLister = serviceInformer.Lister()
 	s.serviceListerSynced = serviceInformer.Informer().HasSynced
 
+	endpointInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
+		UpdateFunc: s.updateEndpoint,
+	})
+
 	nodeInformer.Informer().AddEventHandlerWithResyncPeriod(
 		cache.ResourceEventHandlerFuncs{
 			AddFunc: func(cur interface{}) {
@@ -189,7 +207,6 @@ func New(
 		},
 		time.Duration(0),
 	)
-
 	if err := s.init(); err != nil {
 		return nil, err
 	}
@@ -229,6 +246,7 @@ func (s *Controller) enqueueService(obj interface{}) {
 func (s *Controller) Run(ctx context.Context, workers int) {
 	defer runtime.HandleCrash()
 	defer s.queue.ShutDown()
+	defer s.endpointsQueue.ShutDown()
 
 	klog.Info("Starting service controller")
 	defer klog.Info("Shutting down service controller")
@@ -239,10 +257,16 @@ func (s *Controller) Run(ctx context.Context, workers int) {
 
 	for i := 0; i < workers; i++ {
 		go wait.UntilWithContext(ctx, s.worker, time.Second)
+		if s.directPodIP {
+			go wait.UntilWithContext(ctx, s.endpointWorker, time.Second)
+		}
 	}
 
-	go s.nodeSyncLoop(ctx, workers)
-	go wait.Until(s.triggerNodeSync, nodeSyncPeriod, ctx.Done())
+	// sync node only if LB svc in nodePort mode
+	if !s.directPodIP {
+		go s.nodeSyncLoop(ctx, workers)
+		go wait.Until(s.triggerNodeSync, nodeSyncPeriod, ctx.Done())
+	}
 
 	<-ctx.Done()
 }
@@ -284,6 +308,50 @@ func (s *Controller) worker(ctx context.Context) {
 	}
 }
 
+// endpointWorker runs a worker thread that just dequeues items, processes them, and marks them done.
+// It enforces that the syncHandler is never invoked concurrently with the same key.
+func (s *Controller) endpointWorker(ctx context.Context) {
+	for s.processNextEndpointWorkItem(ctx) {
+	}
+}
+
+func (s *Controller) processNextEndpointWorkItem(ctx context.Context) bool {
+	key, quit := s.endpointsQueue.Get()
+	if quit {
+		return false
+	}
+	defer s.endpointsQueue.Done(key)
+
+	err := s.updateService(ctx, key.(string))
+	if err == nil {
+		s.endpointsQueue.Forget(key)
+		return true
+	}
+
+	runtime.HandleError(fmt.Errorf("error updating endpoints for service %v (will retry): %v", key, err))
+	s.endpointsQueue.AddRateLimited(key)
+	return true
+}
+
+// updateService will call updateLoadBalancer to update pool members
+func (s *Controller) updateService(ctx context.Context, key string) error {
+
+	namespace, name, err := cache.SplitMetaNamespaceKey(key)
+	if err != nil {
+		return err
+	}
+	service, err := s.serviceLister.Services(namespace).Get(name)
+	if err != nil {
+		klog.Errorf("Failed to get service %s with error %v", name, err)
+		return err
+	}
+	klog.Infof("Update service %s/%s which in direct-pod-ip mode triggered by endpoints changed ", namespace, name)
+	var emptyNodes [] *v1.Node
+	err = s.balancer.UpdateLoadBalancer(ctx, s.clusterName, service, emptyNodes)
+
+	return err
+}
+
 // nodeSyncLoop takes nodeSync signal and triggers nodeSync
 func (s *Controller) nodeSyncLoop(ctx context.Context, workers int) {
 	klog.V(4).Info("nodeSyncLoop Started")
@@ -604,6 +672,47 @@ func (s *Controller) needsUpdate(oldService *v1.Service, newService *v1.Service)
 	return false
 }
 
+// updateEndpoint
+func (s *Controller) updateEndpoint(old, new interface{}){
+	oldep := old.(*v1.Endpoints)
+	newep := new.(*v1.Endpoints)
+
+	// service of the endpoint
+	name := newep.Name
+	namespace := newep.Namespace
+
+	service, err := s.serviceLister.Services(namespace).Get(name)
+
+	if errors.IsNotFound(err) {
+		// for some endpoints without service that used for leader election, such as endpoints in kubevirt namespace
+		// the resourceVersion keep increasing and endpoints keep updating
+		klog.V(4).Infof("Failed to get service %s with error %v, do not need update service", name, err)
+		return
+	} else if err != nil {
+		klog.Errorf("Failed to get service %s with error %v", name, err)
+		return
+	}
+
+	if service.Spec.Type == v1.ServiceTypeLoadBalancer {
+		if s.directPodIP {
+			if !reflect.DeepEqual(oldep.Subsets, newep.Subsets) {
+				klog.Infof("Endpoint %s changed, update the service in direct-pod-ip mode", name)
+				s.enqueueEndpoint(new)
+			}
+		}
+	}
+}
+
+// obj could be an *v1.Endpoints, or a DeletionFinalStateUnknown marker item.
+func (s *Controller) enqueueEndpoint(obj interface{}) {
+	key, err := cache.DeletionHandlingMetaNamespaceKeyFunc(obj)
+	if err != nil {
+		runtime.HandleError(fmt.Errorf("couldn't get key for object %#v: %v", obj, err))
+		return
+	}
+	s.endpointsQueue.Add(key)
+}
+
 func getPortsForLB(service *v1.Service) []*v1.ServicePort {
 	ports := []*v1.ServicePort{}
 	for i := range service.Spec.Ports {
@@ -770,6 +879,11 @@ func (s *Controller) nodeSyncService(svc *v1.Service) bool {
 	if svc == nil || !wantsLoadBalancer(svc) {
 		return false
 	}
+	if s.directPodIP {
+		klog.Infof("service %s/%s is direct-pod-ip mode, skip sync nodes", svc.Namespace, svc.Name)
+		return false
+	}
+
 	klog.V(4).Infof("nodeSyncService started for service %s/%s", svc.Namespace, svc.Name)
 	hosts, err := listWithPredicate(s.nodeLister, s.getNodeConditionPredicate())
 	if err != nil {
diff --git a/options/servicecontroller.go b/options/servicecontroller.go
index c63e43a9..abb01428 100644
--- a/options/servicecontroller.go
+++ b/options/servicecontroller.go
@@ -33,6 +33,7 @@ func (o *ServiceControllerOptions) AddFlags(fs *pflag.FlagSet) {
 	}
 
 	fs.Int32Var(&o.ConcurrentServiceSyncs, "concurrent-service-syncs", o.ConcurrentServiceSyncs, "The number of services that are allowed to sync concurrently. Larger number = more responsive service management, but more CPU (and network) load")
+	fs.BoolVar(&o.DirectPodIP, "direct-pod-ip",true, "If 'true' configured, load balancer will use pods ip addresses as pool members instead of nodes addresses.")
 }
 
 // ApplyTo fills up ServiceController config with options.
@@ -42,6 +43,7 @@ func (o *ServiceControllerOptions) ApplyTo(cfg *serviceconfig.ServiceControllerC
 	}
 
 	cfg.ConcurrentServiceSyncs = o.ConcurrentServiceSyncs
+	cfg.DirectPodIP = o.DirectPodIP
 
 	return nil
 }