diff --git a/.buildkite/pipeline.yml b/.buildkite/pipeline.yml
index a644a9023..3bc0c2a66 100644
--- a/.buildkite/pipeline.yml
+++ b/.buildkite/pipeline.yml
@@ -73,14 +73,14 @@ steps:
       - make test-in-docker
     timeout_in_minutes: 10
 
-  - label: ":rotating_light: :running_shirt_with_sash: runtime isolated tests"
+  - label: ":running: runtime isolated tests"
     agents:
       queue: "${BUILDKITE_AGENT_META_DATA_QUEUE:-default}"
       distro: "${BUILDKITE_AGENT_META_DATA_DISTRO}"
       hostname: "${BUILDKITE_AGENT_META_DATA_HOSTNAME}"
     env:
       DOCKER_IMAGE_TAG: "$BUILDKITE_BUILD_NUMBER"
-      NUMBER_OF_VMS: 100
+      NUMBER_OF_VMS: 20
       EXTRAGOARGS: "-v -count=1 -race -timeout=1h"
       FICD_DM_VOLUME_GROUP: fcci-vg
     artifact_paths:
@@ -88,6 +88,23 @@ steps:
     command:
       - make -C runtime integ-test FICD_DM_POOL=build_${BUILDKITE_BUILD_NUMBER}_runtime
 
+  - label: ":weight_lifter: running stress tests"
+    concurrency_group: stress
+    concurrency: 1
+    agents:
+      queue: "${BUILDKITE_AGENT_META_DATA_QUEUE:-default}"
+      distro: "${BUILDKITE_AGENT_META_DATA_DISTRO}"
+      hostname: "${BUILDKITE_AGENT_META_DATA_HOSTNAME}"
+    env:
+      DOCKER_IMAGE_TAG: "$BUILDKITE_BUILD_NUMBER"
+      NUMBER_OF_VMS: 100
+      EXTRAGOARGS: "-v -count=1 -race -timeout=1h -run TestMultipleVMs_Isolated"
+      FICD_DM_VOLUME_GROUP: fcci-vg
+    artifact_paths:
+      - "runtime/logs/*"
+    command:
+      - make -C runtime integ-test FICD_DM_POOL=build_${BUILDKITE_BUILD_NUMBER}_stress
+
   - label: ":rotating_light: :exclamation: example tests"
     agents:
       queue: "${BUILDKITE_AGENT_META_DATA_QUEUE:-default}"
diff --git a/runtime/Makefile b/runtime/Makefile
index 47777018f..05836aa3a 100644
--- a/runtime/Makefile
+++ b/runtime/Makefile
@@ -33,7 +33,7 @@ INTEG_TESTNAMES=$(shell docker run --rm \
 	--entrypoint=/bin/bash \
 	--workdir=/src/runtime \
 	$(FIRECRACKER_CONTAINERD_TEST_IMAGE):$(DOCKER_IMAGE_TAG) \
-	-c "go test -list . | sed '$$d' | grep $(INTEG_TEST_SUFFIX)")
+	-c "go test -list . | sed '$$d' | grep $(INTEG_TEST_SUFFIX)" | grep Multi)
 
 all: runtime
 
diff --git a/runtime/service_integ_test.go b/runtime/service_integ_test.go
index 64685a7eb..ae99b8950 100644
--- a/runtime/service_integ_test.go
+++ b/runtime/service_integ_test.go
@@ -242,14 +242,7 @@ func createTapDevice(ctx context.Context, tapName string) error {
 func TestMultipleVMs_Isolated(t *testing.T) {
 	integtest.Prepare(t)
 
-	// This test starts multiple VMs and some may hit firecracker-containerd's
-	// default timeout. So overriding the timeout to wait longer.
-	// One hour should be enough to start a VM, regardless of the load of
-	// the underlying host.
-	const createVMTimeout = time.Hour
-
-	netns, err := ns.GetCurrentNS()
-	require.NoError(t, err, "failed to get a namespace")
+	var err error
 
 	// numberOfVmsEnvName = NUMBER_OF_VMS ENV and is configurable from buildkite
 	numberOfVms := defaultNumberOfVms
@@ -257,7 +250,36 @@ func TestMultipleVMs_Isolated(t *testing.T) {
 		numberOfVms, err = strconv.Atoi(str)
 		require.NoError(t, err, "failed to get NUMBER_OF_VMS env")
 	}
-	t.Logf("TestMultipleVMs_Isolated: will run %d vm's", numberOfVms)
+	t.Logf("TestMultipleVMs_Isolated: will run up to %d VMs", numberOfVms)
+
+	// We should be able to run 10 VMs without any issues.
+	if numberOfVms <= 10 {
+		testMultipleVMs(t, 10)
+		return
+	}
+
+	// We have issues running 100 VMs (see #581).
+	// Incrementally increase the number of VMs to find the breaking point.
+	for n := 0; n <= numberOfVms; n += 10 {
+		success := t.Run(fmt.Sprintf("VMs=%d", n), func(t *testing.T) {
+			testMultipleVMs(t, n)
+		})
+		if !success {
+			// If running N VMs doesn't work, no point to go further.
+			return
+		}
+	}
+}
+
+func testMultipleVMs(t *testing.T, count int) {
+	// This test starts multiple VMs and some may hit firecracker-containerd's
+	// default timeout. So overriding the timeout to wait longer.
+	// One hour should be enough to start a VM, regardless of the load of
+	// the underlying host.
+	const createVMTimeout = 1 * time.Hour
+
+	netns, err := ns.GetCurrentNS()
+	require.NoError(t, err, "failed to get a namespace")
 
 	tapPrefix := os.Getenv(tapPrefixEnvName)
 
@@ -303,7 +325,7 @@ func TestMultipleVMs_Isolated(t *testing.T) {
 	// container ends up in the right VM by assigning each VM a network device with a unique mac address and having each container
 	// print the mac address it sees inside its VM.
 	vmEg, vmEgCtx := errgroup.WithContext(testCtx)
-	for i := 0; i < numberOfVms; i++ {
+	for i := 0; i < count; i++ {
 		caseTypeNumber := i % len(cases)
 		vmID := i
 		c := cases[caseTypeNumber]
@@ -349,6 +371,7 @@ func TestMultipleVMs_Isolated(t *testing.T) {
 			if err != nil {
 				return err
 			}
+			defer fcClient.Close()
 
 			resp, createVMErr := fcClient.CreateVM(ctx, req)
 			if createVMErr != nil {
@@ -423,12 +446,10 @@ func TestMultipleVMs_Isolated(t *testing.T) {
 			if err != nil {
 				return fmt.Errorf("unexpected error from the containers in VM %d: %w", vmID, err)
 			}
+			t.Logf("all containers in VM %d are stopped", vmID)
 
 			_, err = fcClient.StopVM(ctx, &proto.StopVMRequest{VMID: strconv.Itoa(vmID), TimeoutSeconds: 5})
-			if err != nil {
-				return err
-			}
-			return nil
+			return err
 		}
 
 		vmEg.Go(func() error {
@@ -478,7 +499,7 @@ func testMultipleExecs(
 	if err != nil {
 		return err
 	}
-	defer newContainer.Delete(ctx)
+	defer newContainer.Delete(ctx, containerd.WithSnapshotCleanup)
 
 	var taskStdout bytes.Buffer
 	var taskStderr bytes.Buffer