diff --git a/test/e2e/kubeproxy.go b/test/e2e/kubeproxy.go
index 9b9b5afc435..2949d009a22 100644
--- a/test/e2e/kubeproxy.go
+++ b/test/e2e/kubeproxy.go
@@ -31,6 +31,7 @@ import (
 	"k8s.io/kubernetes/pkg/api/latest"
 	"k8s.io/kubernetes/pkg/api/unversioned"
 	client "k8s.io/kubernetes/pkg/client/unversioned"
+	"k8s.io/kubernetes/pkg/labels"
 	"k8s.io/kubernetes/pkg/util"
 	"k8s.io/kubernetes/pkg/util/wait"
 )
@@ -471,7 +472,18 @@ func (config *KubeProxyTestConfig) deleteNetProxyPod() {
 	pod := config.endpointPods[0]
 	config.getPodClient().Delete(pod.Name, nil)
 	config.endpointPods = config.endpointPods[1:]
-	time.Sleep(5 * time.Second) // wait for kube-proxy to catch up with the pod being deleted.
+	// wait for pod being deleted.
+	err := waitForPodToDisappear(config.f.Client, config.f.Namespace.Name, pod.Name, labels.Everything(), time.Second, util.ForeverTestTimeout)
+	if err != nil {
+		Failf("Failed to delete %s pod: %v", pod.Name, err)
+	}
+	// wait for endpoint being removed.
+	err = waitForServiceEndpointsNum(config.f.Client, config.f.Namespace.Name, nodePortServiceName, len(config.endpointPods), time.Second, util.ForeverTestTimeout)
+	if err != nil {
+		Failf("Failed to remove endpoint from service: %s", nodePortServiceName)
+	}
+	// wait for kube-proxy to catch up with the pod being deleted.
+	time.Sleep(5 * time.Second)
 }
 
 func (config *KubeProxyTestConfig) createPod(pod *api.Pod) *api.Pod {
diff --git a/test/e2e/util.go b/test/e2e/util.go
index 5745cb08381..d24a6973ff7 100644
--- a/test/e2e/util.go
+++ b/test/e2e/util.go
@@ -641,11 +641,8 @@ func waitForRCPodOnNode(c *client.Client, ns, rcName, node string) (*api.Pod, er
 	return p, err
 }
 
-// waitForRCPodToDisappear returns nil if the pod from the given replication controller (described by rcName) no longer exists.
-// In case of failure or too long waiting time, an error is returned.
-func waitForRCPodToDisappear(c *client.Client, ns, rcName, podName string) error {
-	label := labels.SelectorFromSet(labels.Set(map[string]string{"name": rcName}))
-	return wait.Poll(20*time.Second, 5*time.Minute, func() (bool, error) {
+func waitForPodToDisappear(c *client.Client, ns, podName string, label labels.Selector, interval, timeout time.Duration) error {
+	return wait.Poll(interval, timeout, func() (bool, error) {
 		Logf("Waiting for pod %s to disappear", podName)
 		pods, err := c.Pods(ns).List(label, fields.Everything())
 		if err != nil {
@@ -666,6 +663,13 @@ func waitForRCPodToDisappear(c *client.Client, ns, rcName, podName string) error
 	})
 }
 
+// waitForRCPodToDisappear returns nil if the pod from the given replication controller (described by rcName) no longer exists.
+// In case of failure or too long waiting time, an error is returned.
+func waitForRCPodToDisappear(c *client.Client, ns, rcName, podName string) error {
+	label := labels.SelectorFromSet(labels.Set(map[string]string{"name": rcName}))
+	return waitForPodToDisappear(c, ns, podName, label, 20*time.Second, 5*time.Minute)
+}
+
 // waitForService waits until the service appears (exist == true), or disappears (exist == false)
 func waitForService(c *client.Client, namespace, name string, exist bool, interval, timeout time.Duration) error {
 	err := wait.Poll(interval, timeout, func() (bool, error) {
@@ -695,6 +699,32 @@ func waitForService(c *client.Client, namespace, name string, exist bool, interv
 	return nil
 }
 
+//waitForServiceEndpointsNum waits until the amount of endpoints that implement service to expectNum.
+func waitForServiceEndpointsNum(c *client.Client, namespace, serviceName string, expectNum int, interval, timeout time.Duration) error {
+	return wait.Poll(interval, timeout, func() (bool, error) {
+		Logf("Waiting for amount of service:%s endpoints to %d", serviceName, expectNum)
+		list, err := c.Endpoints(namespace).List(labels.Everything())
+		if err != nil {
+			return false, err
+		}
+
+		for _, e := range list.Items {
+			if e.Name == serviceName && countEndpointsNum(&e) == expectNum {
+				return true, nil
+			}
+		}
+		return false, nil
+	})
+}
+
+func countEndpointsNum(e *api.Endpoints) int {
+	num := 0
+	for _, sub := range e.Subsets {
+		num += len(sub.Addresses)
+	}
+	return num
+}
+
 // waitForReplicationController waits until the RC appears (exist == true), or disappears (exist == false)
 func waitForReplicationController(c *client.Client, namespace, name string, exist bool, interval, timeout time.Duration) error {
 	err := wait.Poll(interval, timeout, func() (bool, error) {