diff --git a/test/e2e_node/density_test.go b/test/e2e_node/density_test.go
index 5bff7ff8b6..835d619d60 100644
--- a/test/e2e_node/density_test.go
+++ b/test/e2e_node/density_test.go
@@ -19,7 +19,6 @@ limitations under the License.
 package e2e_node
 
 import (
-	"errors"
 	"fmt"
 	"sort"
 	"strconv"
@@ -27,7 +26,6 @@ import (
 	"time"
 
 	"k8s.io/kubernetes/pkg/api"
-	apierrors "k8s.io/kubernetes/pkg/api/errors"
 	"k8s.io/kubernetes/pkg/api/unversioned"
 	"k8s.io/kubernetes/pkg/client/cache"
 	controllerframework "k8s.io/kubernetes/pkg/controller/framework"
@@ -67,15 +65,12 @@ var _ = framework.KubeDescribe("Density [Serial] [Slow]", func() {
 		ns = f.Namespace.Name
 		nodeName = framework.TestContext.NodeName
 		// Start a standalone cadvisor pod using 'createSync', the pod is running when it returns
-		createCadvisorPod(f)
+		f.PodClient().CreateSync(getCadvisorPod())
 		// Resource collector monitors fine-grain CPU/memory usage by a standalone Cadvisor with
 		// 1s housingkeeping interval
 		rc = NewResourceCollector(containerStatsPollingPeriod)
 	})
 
-	AfterEach(func() {
-	})
-
 	Context("create a batch of pods", func() {
 		// TODO(coufon): the values are generous, set more precise limits with benchmark data
 		// and add more tests
@@ -296,6 +291,8 @@ func runDensityBatchTest(f *framework.Framework, rc *ResourceCollector, testArg
 	time.Sleep(sleepBeforeCreatePods)
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(pods, getCadvisorPod()))
 	defer rc.Stop()
 
 	By("Creating a batch of pods")
@@ -371,6 +368,8 @@ func runDensitySeqTest(f *framework.Framework, rc *ResourceCollector, testArg de
 	time.Sleep(sleepBeforeCreatePods)
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(bgPods, append(testPods, getCadvisorPod())...))
 	defer rc.Stop()
 
 	// Create pods sequentially (back-to-back). e2eLags have been sorted.
@@ -394,16 +393,6 @@ func createBatchPodWithRateControl(f *framework.Framework, pods []*api.Pod, inte
 	return createTimes
 }
 
-// checkPodDeleted checks whether a pod has been successfully deleted
-func checkPodDeleted(f *framework.Framework, podName string) error {
-	ns := f.Namespace.Name
-	_, err := f.Client.Pods(ns).Get(podName)
-	if apierrors.IsNotFound(err) {
-		return nil
-	}
-	return errors.New("Pod Not Deleted")
-}
-
 // getPodStartLatency gets prometheus metric 'pod start latency' from kubelet
 func getPodStartLatency(node string) (framework.KubeletLatencyMetrics, error) {
 	latencyMetrics := framework.KubeletLatencyMetrics{}
diff --git a/test/e2e_node/resource_collector.go b/test/e2e_node/resource_collector.go
index 7928583174..f7498c6e37 100644
--- a/test/e2e_node/resource_collector.go
+++ b/test/e2e_node/resource_collector.go
@@ -292,8 +292,8 @@ func formatCPUSummary(summary framework.ContainersCPUSummary) string {
 }
 
 // createCadvisorPod creates a standalone cadvisor pod for fine-grain resource monitoring.
-func createCadvisorPod(f *framework.Framework) {
-	f.PodClient().CreateSync(&api.Pod{
+func getCadvisorPod() *api.Pod {
+	return &api.Pod{
 		ObjectMeta: api.ObjectMeta{
 			Name: cadvisorPodName,
 		},
@@ -363,24 +363,22 @@ func createCadvisorPod(f *framework.Framework) {
 				},
 			},
 		},
-	})
+	}
 }
 
-// deleteBatchPod deletes a batch of pods (synchronous).
-func deleteBatchPod(f *framework.Framework, pods []*api.Pod) {
-	ns := f.Namespace.Name
+// deletePodsSync deletes a list of pods and block until pods disappear.
+func deletePodsSync(f *framework.Framework, pods []*api.Pod) {
 	var wg sync.WaitGroup
 	for _, pod := range pods {
 		wg.Add(1)
 		go func(pod *api.Pod) {
 			defer wg.Done()
 
-			err := f.Client.Pods(ns).Delete(pod.ObjectMeta.Name, api.NewDeleteOptions(30))
+			err := f.PodClient().Delete(pod.ObjectMeta.Name, api.NewDeleteOptions(30))
 			Expect(err).NotTo(HaveOccurred())
 
-			Expect(framework.WaitForPodToDisappear(f.Client, ns, pod.ObjectMeta.Name, labels.Everything(),
-				30*time.Second, 10*time.Minute)).
-				NotTo(HaveOccurred())
+			Expect(framework.WaitForPodToDisappear(f.Client, f.Namespace.Name, pod.ObjectMeta.Name, labels.Everything(),
+				30*time.Second, 10*time.Minute)).NotTo(HaveOccurred())
 		}(pod)
 	}
 	wg.Wait()
diff --git a/test/e2e_node/resource_usage_test.go b/test/e2e_node/resource_usage_test.go
index 4dd7f899af..ebb2858926 100644
--- a/test/e2e_node/resource_usage_test.go
+++ b/test/e2e_node/resource_usage_test.go
@@ -52,7 +52,7 @@ var _ = framework.KubeDescribe("Resource-usage [Serial] [Slow]", func() {
 		// The Cadvsior of Kubelet has a housekeeping interval of 10s, which is too long to
 		// show the resource usage spikes. But changing its interval increases the overhead
 		// of kubelet. Hence we use a Cadvisor pod.
-		createCadvisorPod(f)
+		f.PodClient().CreateSync(getCadvisorPod())
 		rc = NewResourceCollector(containerStatsPollingPeriod)
 	})
 
@@ -135,12 +135,14 @@ func runResourceUsageTest(f *framework.Framework, rc *ResourceCollector, testArg
 		// sleep for an interval here to measure steady data
 		sleepAfterCreatePods = 10 * time.Second
 	)
+	pods := newTestPods(testArg.podsNr, ImageRegistry[pauseImage], "test_pod")
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(pods, getCadvisorPod()))
 	defer rc.Stop()
 
 	By("Creating a batch of Pods")
-	pods := newTestPods(testArg.podsNr, ImageRegistry[pauseImage], "test_pod")
 	f.PodClient().CreateBatch(pods)
 
 	// wait for a while to let the node be steady