diff --git a/pkg/kubelet/eviction/helpers.go b/pkg/kubelet/eviction/helpers.go
index ea6911240d2..2709153cd8c 100644
--- a/pkg/kubelet/eviction/helpers.go
+++ b/pkg/kubelet/eviction/helpers.go
@@ -695,3 +695,7 @@ func buildResourceToRankFunc(withImageFs bool) map[api.ResourceName]rankFunc {
 	}
 	return resourceToRankFunc
 }
+
+func PodIsEvicted(podStatus api.PodStatus) bool {
+	return podStatus.Phase == api.PodFailed && podStatus.Reason == reason
+}
diff --git a/pkg/kubelet/kubelet.go b/pkg/kubelet/kubelet.go
index 03c5c25191d..7d8a491261b 100644
--- a/pkg/kubelet/kubelet.go
+++ b/pkg/kubelet/kubelet.go
@@ -2188,11 +2188,10 @@ func (kl *Kubelet) syncLoopIteration(configCh <-chan kubetypes.PodUpdate, handle
 			glog.V(2).Infof("SyncLoop (PLEG): %q, event: %#v", format.Pod(pod), e)
 			handler.HandlePodSyncs([]*api.Pod{pod})
 		}
+
 		if e.Type == pleg.ContainerDied {
-			if podStatus, err := kl.podCache.Get(e.ID); err == nil {
-				if containerID, ok := e.Data.(string); ok {
-					kl.containerDeletor.deleteContainersInPod(containerID, podStatus)
-				}
+			if containerID, ok := e.Data.(string); ok {
+				kl.cleanUpContainersInPod(e.ID, containerID)
 			}
 		}
 	case <-syncCh:
@@ -2922,6 +2921,16 @@ func (kl *Kubelet) ListenAndServeReadOnly(address net.IP, port uint) {
 	server.ListenAndServeKubeletReadOnlyServer(kl, kl.resourceAnalyzer, address, port, kl.containerRuntime)
 }
 
+// Delete the eligible dead container instances in a pod. Depending on the configuration, the latest dead containers may be kept around.
+func (kl *Kubelet) cleanUpContainersInPod(podId types.UID, exitedContainerID string) {
+	if podStatus, err := kl.podCache.Get(podId); err == nil {
+		if status, ok := kl.statusManager.GetPodStatus(podId); ok {
+			// If a pod is evicted, we can delete all the dead containers.
+			kl.containerDeletor.deleteContainersInPod(exitedContainerID, podStatus, eviction.PodIsEvicted(status))
+		}
+	}
+}
+
 // isSyncPodWorthy filters out events that are not worthy of pod syncing
 func isSyncPodWorthy(event *pleg.PodLifecycleEvent) bool {
 	// ContatnerRemoved doesn't affect pod state
diff --git a/pkg/kubelet/pod_container_deletor.go b/pkg/kubelet/pod_container_deletor.go
index d61228947d3..122170409a3 100644
--- a/pkg/kubelet/pod_container_deletor.go
+++ b/pkg/kubelet/pod_container_deletor.go
@@ -59,7 +59,7 @@ func newPodContainerDeletor(runtime kubecontainer.Runtime, containersToKeep int)
 }
 
 // getContainersToDeleteInPod returns the exited containers in a pod whose name matches the name inferred from exitedContainerID, ordered by the creation time from the latest to the earliest.
-func (p *podContainerDeletor) getContainersToDeleteInPod(exitedContainerID string, podStatus *kubecontainer.PodStatus) containerStatusbyCreatedList {
+func getContainersToDeleteInPod(exitedContainerID string, podStatus *kubecontainer.PodStatus, containersToKeep int) containerStatusbyCreatedList {
 	var matchedContainer *kubecontainer.ContainerStatus
 	var exitedContainers []*kubecontainer.ContainerStatus
 	// Find all exited containers in the pod
@@ -84,17 +84,21 @@ func (p *podContainerDeletor) getContainersToDeleteInPod(exitedContainerID strin
 			candidates = append(candidates, containerStatus)
 		}
 	}
-	if len(candidates) <= p.containersToKeep {
+
+	if len(candidates) <= containersToKeep {
 		return containerStatusbyCreatedList{}
 	}
-
 	sort.Sort(candidates)
-	return candidates[p.containersToKeep:]
+	return candidates[containersToKeep:]
 }
 
 // deleteContainersInPod issues container deletion requests for containers selected by getContainersToDeleteInPod.
-func (p *podContainerDeletor) deleteContainersInPod(exitedContainerID string, podStatus *kubecontainer.PodStatus) {
-	for _, candidate := range p.getContainersToDeleteInPod(exitedContainerID, podStatus) {
+func (p *podContainerDeletor) deleteContainersInPod(exitedContainerID string, podStatus *kubecontainer.PodStatus, removeAll bool) {
+	containersToKeep := p.containersToKeep
+	if removeAll {
+		containersToKeep = 0
+	}
+	for _, candidate := range getContainersToDeleteInPod(exitedContainerID, podStatus, containersToKeep) {
 		select {
 		case p.worker <- candidate.ID:
 		default:
diff --git a/pkg/kubelet/pod_container_deletor_test.go b/pkg/kubelet/pod_container_deletor_test.go
index 52a1b8f730c..51797d14c3e 100644
--- a/pkg/kubelet/pod_container_deletor_test.go
+++ b/pkg/kubelet/pod_container_deletor_test.go
@@ -22,7 +22,6 @@ import (
 	"time"
 
 	kubecontainer "k8s.io/kubernetes/pkg/kubelet/container"
-	containertest "k8s.io/kubernetes/pkg/kubelet/container/testing"
 )
 
 func testGetContainersToDeleteInPod(t *testing.T) {
@@ -61,9 +60,28 @@ func testGetContainersToDeleteInPod(t *testing.T) {
 		},
 	}
 
-	expectedCandidates := []*kubecontainer.ContainerStatus{pod.ContainerStatuses[2], pod.ContainerStatuses[1]}
-	candidates := newPodContainerDeletor(&containertest.FakeRuntime{}, 1).getContainersToDeleteInPod("2", &pod)
-	if !reflect.DeepEqual(candidates, expectedCandidates) {
-		t.Errorf("expected %v got %v", expectedCandidates, candidates)
+	testCases := []struct {
+		containersToKeep           int
+		expectedContainersToDelete []*kubecontainer.ContainerStatus
+	}{
+		{
+			0,
+			[]*kubecontainer.ContainerStatus{pod.ContainerStatuses[3], pod.ContainerStatuses[2], pod.ContainerStatuses[1]},
+		},
+		{
+			1,
+			[]*kubecontainer.ContainerStatus{pod.ContainerStatuses[2], pod.ContainerStatuses[1]},
+		},
+		{
+			2,
+			[]*kubecontainer.ContainerStatus{pod.ContainerStatuses[1]},
+		},
+	}
+
+	for _, test := range testCases {
+		candidates := getContainersToDeleteInPod("4", &pod, test.containersToKeep)
+		if !reflect.DeepEqual(getContainersToDeleteInPod("4", &pod, test.containersToKeep), test.expectedContainersToDelete) {
+			t.Errorf("expected %v got %v", test.expectedContainersToDelete, candidates)
+		}
 	}
 }