diff --git a/test/e2e/framework/util.go b/test/e2e/framework/util.go
index 0071406da7c..9852c9d3870 100644
--- a/test/e2e/framework/util.go
+++ b/test/e2e/framework/util.go
@@ -2471,12 +2471,12 @@ func (config *RCConfig) start() error {
 }
 
 // Simplified version of RunRC, that does not create RC, but creates plain Pods.
-// optionally waits for pods to start running (if waitForRunning == true)
+// Optionally waits for pods to start running (if waitForRunning == true).
+// The number of replicas must be non-zero.
 func StartPods(c *client.Client, replicas int, namespace string, podNamePrefix string, pod api.Pod, waitForRunning bool) {
 	// no pod to start
 	if replicas < 1 {
-		Logf("No pod to start, skipping...")
-		return
+		panic("StartPods: number of replicas must be non-zero")
 	}
 	startPodsID := string(util.NewUUID()) // So that we can label and find them
 	for i := 0; i < replicas; i++ {
diff --git a/test/e2e/scheduler_predicates.go b/test/e2e/scheduler_predicates.go
index a822c4be6e4..0849cb1fc67 100644
--- a/test/e2e/scheduler_predicates.go
+++ b/test/e2e/scheduler_predicates.go
@@ -235,24 +235,29 @@ var _ = framework.KubeDescribe("SchedulerPredicates [Serial]", func() {
 
 		By(fmt.Sprintf("Starting additional %v Pods to fully saturate the cluster max pods and trying to start another one", podsNeededForSaturation))
 
-		framework.StartPods(c, podsNeededForSaturation, ns, "maxp", api.Pod{
-			TypeMeta: unversioned.TypeMeta{
-				Kind: "Pod",
-			},
-			ObjectMeta: api.ObjectMeta{
-				Name:   "",
-				Labels: map[string]string{"name": ""},
-			},
-			Spec: api.PodSpec{
-				Containers: []api.Container{
-					{
-						Name:  "",
-						Image: framework.GetPauseImageName(f.Client),
+		// As the pods are distributed randomly among nodes,
+		// it can easily happen that all nodes are satured
+		// and there is no need to create additional pods.
+		// StartPods requires at least one pod to replicate.
+		if podsNeededForSaturation > 0 {
+			framework.StartPods(c, podsNeededForSaturation, ns, "maxp", api.Pod{
+				TypeMeta: unversioned.TypeMeta{
+					Kind: "Pod",
+				},
+				ObjectMeta: api.ObjectMeta{
+					Name:   "",
+					Labels: map[string]string{"name": ""},
+				},
+				Spec: api.PodSpec{
+					Containers: []api.Container{
+						{
+							Name:  "",
+							Image: framework.GetPauseImageName(f.Client),
+						},
 					},
 				},
-			},
-		}, true)
-
+			}, true)
+		}
 		podName := "additional-pod"
 		_, err := c.Pods(ns).Create(&api.Pod{
 			TypeMeta: unversioned.TypeMeta{
@@ -312,32 +317,37 @@ var _ = framework.KubeDescribe("SchedulerPredicates [Serial]", func() {
 
 		By(fmt.Sprintf("Starting additional %v Pods to fully saturate the cluster CPU and trying to start another one", podsNeededForSaturation))
 
-		framework.StartPods(c, podsNeededForSaturation, ns, "overcommit", api.Pod{
-			TypeMeta: unversioned.TypeMeta{
-				Kind: "Pod",
-			},
-			ObjectMeta: api.ObjectMeta{
-				Name:   "",
-				Labels: map[string]string{"name": ""},
-			},
-			Spec: api.PodSpec{
-				Containers: []api.Container{
-					{
-						Name:  "",
-						Image: framework.GetPauseImageName(f.Client),
-						Resources: api.ResourceRequirements{
-							Limits: api.ResourceList{
-								"cpu": *resource.NewMilliQuantity(milliCpuPerPod, "DecimalSI"),
-							},
-							Requests: api.ResourceList{
-								"cpu": *resource.NewMilliQuantity(milliCpuPerPod, "DecimalSI"),
+		// As the pods are distributed randomly among nodes,
+		// it can easily happen that all nodes are saturated
+		// and there is no need to create additional pods.
+		// StartPods requires at least one pod to replicate.
+		if podsNeededForSaturation > 0 {
+			framework.StartPods(c, podsNeededForSaturation, ns, "overcommit", api.Pod{
+				TypeMeta: unversioned.TypeMeta{
+					Kind: "Pod",
+				},
+				ObjectMeta: api.ObjectMeta{
+					Name:   "",
+					Labels: map[string]string{"name": ""},
+				},
+				Spec: api.PodSpec{
+					Containers: []api.Container{
+						{
+							Name:  "",
+							Image: framework.GetPauseImageName(f.Client),
+							Resources: api.ResourceRequirements{
+								Limits: api.ResourceList{
+									"cpu": *resource.NewMilliQuantity(milliCpuPerPod, "DecimalSI"),
+								},
+								Requests: api.ResourceList{
+									"cpu": *resource.NewMilliQuantity(milliCpuPerPod, "DecimalSI"),
+								},
 							},
 						},
 					},
 				},
-			},
-		}, true)
-
+			}, true)
+		}
 		podName := "additional-pod"
 		_, err = c.Pods(ns).Create(&api.Pod{
 			TypeMeta: unversioned.TypeMeta{
diff --git a/test/e2e/ubernetes_lite.go b/test/e2e/ubernetes_lite.go
index c3368c79706..5584a507944 100644
--- a/test/e2e/ubernetes_lite.go
+++ b/test/e2e/ubernetes_lite.go
@@ -93,6 +93,11 @@ func SpreadServiceOrFail(f *framework.Framework, replicaCount int, image string)
 			},
 		},
 	}
+
+	// Caution: StartPods requires at least one pod to replicate.
+	// Based on the callers, replicas is always positive number: zoneCount >= 0 implies (2*zoneCount)+1 > 0.
+	// Thus, no need to test for it. Once the precondition changes to zero number of replicas,
+	// test for replicaCount > 0. Otherwise, StartPods panics.
 	framework.StartPods(f.Client, replicaCount, f.Namespace.Name, serviceName, *podSpec, false)
 
 	// Wait for all of them to be scheduled