node: metrics: add resource alignment metrics

In order to improve the observability of the resource management
in kubelet, cpu allocation and NUMA alignment, we add more metrics
to report if resource alignment is in effect.

The more precise reporting would probably be using pod status,
but this would require more invasive and riskier changes,
and possibly extra interactions to the APIServer.

We start adding metrics to report if containers got their
compute resources aligned.
If metrics are growing, the assingment is working as expected;
If metrics stay consistent, perhaps at zero, no resource
alignment is done.

Extra fixes brought by this work
- retroactively add labels for existing tests
- running metrics test demands precision accounting to avoid flakes;
  ensure the node state is restored pristine between each test, to
  minimize the aforementioned risk of flakes.
- The test pod command line was wrong, with this the pod could not
  reach Running state. That gone unnoticed so far because
  no test using this utility function actually needed a pod
  in running state.

Signed-off-by: Francesco Romani <fromani@redhat.com>

pkg/kubelet/cm/cpumanager/policy_static.go

@@ -312,7 +312,13 @@ func (p *staticPolicy) Allocate(s state.State, pod *v1.Pod, container *v1.Contai
 	defer func() {
 		if rerr != nil {
 			metrics.CPUManagerPinningErrorsTotal.Inc()
+			return
 		}
+		if !p.options.FullPhysicalCPUsOnly {
+			// increment only if we know we allocate aligned resources
+			return
+		}
+		metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedPhysicalCPU).Inc()
 	}()
 
 	if p.options.FullPhysicalCPUsOnly {

pkg/kubelet/cm/topologymanager/policy.go

@@ -30,6 +30,14 @@ type Policy interface {
 	Merge(providersHints []map[string][]TopologyHint) (TopologyHint, bool)
 }
 
+// IsAlignmentGuaranteed return true if the given policy guarantees that either
+// the compute resources will be allocated within a NUMA boundary, or the allocation will fail at all.
+func IsAlignmentGuaranteed(p Policy) bool {
+	// We are abusing the name, but atm this matches almost 1:1 the policy name
+	// so we are not adding new fields for now.
+	return p.Name() == PolicySingleNumaNode
+}
+
 // Merge a TopologyHints permutation to a single hint by performing a bitwise-AND
 // of their affinity masks. The hint shall be preferred if all hits in the permutation
 // are preferred.

pkg/kubelet/cm/topologymanager/scope_container.go

@@ -61,6 +61,10 @@ func (s *containerScope) Admit(pod *v1.Pod) lifecycle.PodAdmitResult {
 			metrics.TopologyManagerAdmissionErrorsTotal.Inc()
 			return admission.GetPodAdmitResult(err)
 		}
+
+		if IsAlignmentGuaranteed(s.policy) {
+			metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedNUMANode).Inc()
+		}
 	}
 	return admission.GetPodAdmitResult(nil)
 }

pkg/kubelet/cm/topologymanager/scope_pod.go

@@ -62,6 +62,10 @@ func (s *podScope) Admit(pod *v1.Pod) lifecycle.PodAdmitResult {
 			return admission.GetPodAdmitResult(err)
 		}
 	}
+	if IsAlignmentGuaranteed(s.policy) {
+		// increment only if we know we allocate aligned resources.
+		metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopePod, metrics.AlignedNUMANode).Inc()
+	}
 	return admission.GetPodAdmitResult(nil)
 }
 

pkg/kubelet/metrics/metrics.go

@@ -127,10 +127,21 @@ const (
 	// Metric for tracking garbage collected images
 	ImageGarbageCollectedTotalKey = "image_garbage_collected_total"
 
+	// Metric for tracking aligment of compute resources
+	ContainerAlignedComputeResourcesNameKey          = "container_aligned_compute_resources_count"
+	ContainerAlignedComputeResourcesScopeLabelKey    = "scope"
+	ContainerAlignedComputeResourcesBoundaryLabelKey = "boundary"
+
 	// Values used in metric labels
 	Container          = "container"
 	InitContainer      = "init_container"
 	EphemeralContainer = "ephemeral_container"
+
+	AlignScopePod       = "pod"
+	AlignScopeContainer = "container"
+
+	AlignedPhysicalCPU = "physical_cpu"
+	AlignedNUMANode    = "numa_node"
 )
 
 type imageSizeBucket struct {
@@ -762,6 +773,16 @@ var (
 		},
 	)
 
+	ContainerAlignedComputeResources = metrics.NewCounterVec(
+		&metrics.CounterOpts{
+			Subsystem:      KubeletSubsystem,
+			Name:           ContainerAlignedComputeResourcesNameKey,
+			Help:           "Cumulative number of aligned compute resources allocated to containers by alignment type.",
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{ContainerAlignedComputeResourcesScopeLabelKey, ContainerAlignedComputeResourcesBoundaryLabelKey},
+	)
+
 	// MemoryManagerPinningRequestTotal tracks the number of times the pod spec required the memory manager to pin memory pages
 	MemoryManagerPinningRequestTotal = metrics.NewCounter(
 		&metrics.CounterOpts{
@@ -985,6 +1006,7 @@ func Register(collectors ...metrics.StableCollector) {
 		legacyregistry.MustRegister(RunPodSandboxErrors)
 		legacyregistry.MustRegister(CPUManagerPinningRequestsTotal)
 		legacyregistry.MustRegister(CPUManagerPinningErrorsTotal)
+		legacyregistry.MustRegister(ContainerAlignedComputeResources)
 		if utilfeature.DefaultFeatureGate.Enabled(features.MemoryManager) {
 			legacyregistry.MustRegister(MemoryManagerPinningRequestTotal)
 			legacyregistry.MustRegister(MemoryManagerPinningErrorsTotal)

test/e2e_node/cpu_manager_metrics_test.go

@@ -91,6 +91,7 @@ var _ = SIGDescribe("CPU Manager Metrics", framework.WithSerial(), feature.CPUMa
 		ginkgo.AfterEach(func(ctx context.Context) {
 			if testPod != nil {
 				deletePodSyncByName(ctx, f, testPod.Name)
+				waitForContainerRemoval(ctx, testPod.Spec.Containers[0].Name, testPod.Name, testPod.Namespace)
 			}
 			updateKubeletConfig(ctx, f, oldCfg, true)
 		})
@@ -160,11 +161,32 @@ var _ = SIGDescribe("CPU Manager Metrics", framework.WithSerial(), feature.CPUMa
 			ginkgo.By("Ensuring the metrics match the expectations a few more times")
 			gomega.Consistently(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchResourceMetrics)
 		})
+
+		ginkgo.It("should return updated alignment counters when pod successfully run", func(ctx context.Context) {
+			ginkgo.By("Creating the test pod")
+			testPod = e2epod.NewPodClient(f).Create(ctx, makeGuaranteedCPUExclusiveSleeperPod("count-align-smt-ok", smtLevel))
+
+			// we updated the kubelet config in BeforeEach, so we can assume we start fresh.
+			// being [Serial], we can also assume noone else but us is running pods.
+			ginkgo.By("Checking the cpumanager metrics right after the kubelet restart, with pod should be admitted")
+
+			idFn := makeCustomPairID("scope", "boundary")
+			matchAlignmentMetrics := gstruct.MatchKeys(gstruct.IgnoreExtras, gstruct.Keys{
+				"kubelet_container_aligned_compute_resources_count": gstruct.MatchElements(idFn, gstruct.IgnoreExtras, gstruct.Elements{
+					"container::physical_cpu": timelessSample(1),
+				}),
+			})
+
+			ginkgo.By("Giving the Kubelet time to update the alignment metrics")
+			gomega.Eventually(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchAlignmentMetrics)
+			ginkgo.By("Ensuring the metrics match the expectations about alignment metrics a few more times")
+			gomega.Consistently(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchAlignmentMetrics)
+		})
 	})
 })
 
 func getKubeletMetrics(ctx context.Context) (e2emetrics.KubeletMetrics, error) {
-	ginkgo.By("getting Kubelet metrics from the metrics API")
+	ginkgo.By("Getting Kubelet metrics from the metrics API")
 	return e2emetrics.GrabKubeletMetricsWithoutProxy(ctx, nodeNameOrIP()+":10255", "/metrics")
 }
 
@@ -189,7 +211,7 @@ func makeGuaranteedCPUExclusiveSleeperPod(name string, cpus int) *v1.Pod {
 							v1.ResourceMemory: resource.MustParse("64Mi"),
 						},
 					},
-					Command: []string{"sh", "-c", "sleep", "1d"},
+					Command: []string{"sh", "-c", "sleep 1d"},
 				},
 			},
 		},

test/e2e_node/resource_metrics_test.go

@@ -157,6 +157,13 @@ func containerID(element interface{}) string {
 	return fmt.Sprintf("%s::%s::%s", el.Metric["namespace"], el.Metric["pod"], el.Metric["container"])
 }
 
+func makeCustomPairID(pri, sec string) func(interface{}) string {
+	return func(element interface{}) string {
+		el := element.(*model.Sample)
+		return fmt.Sprintf("%s::%s", el.Metric[model.LabelName(pri)], el.Metric[model.LabelName(sec)])
+	}
+}
+
 func boundedSample(lower, upper interface{}) types.GomegaMatcher {
 	return gstruct.PointTo(gstruct.MatchAllFields(gstruct.Fields{
 		// We already check Metric when matching the Id

test/e2e_node/topology_manager_metrics_test.go

@@ -28,6 +28,7 @@ import (
 	v1 "k8s.io/api/core/v1"
 	kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config"
 	"k8s.io/kubernetes/pkg/kubelet/cm/topologymanager"
+	"k8s.io/kubernetes/pkg/kubelet/metrics"
 	"k8s.io/kubernetes/test/e2e/feature"
 	"k8s.io/kubernetes/test/e2e/framework"
 	e2epod "k8s.io/kubernetes/test/e2e/framework/pod"
@@ -152,6 +153,26 @@ var _ = SIGDescribe("Topology Manager Metrics", framework.WithSerial(), feature.
 			ginkgo.By("Ensuring the metrics match the expectations a few more times")
 			gomega.Consistently(ctx, getKubeletMetrics, 2*time.Minute, 10*time.Second).Should(matchResourceMetrics)
 		})
+
+		ginkgo.It("[alignment] should return updated alignment counters when pod successfully run", func(ctx context.Context) {
+			ginkgo.By("Creating the test pod")
+			testPod = e2epod.NewPodClient(f).Create(ctx, makeGuaranteedCPUExclusiveSleeperPod("count-align-numa-ok", cpusNumPerNUMA))
+
+			// we updated the kubelet config in BeforeEach, so we can assume we start fresh.
+			// being [Serial], we can also assume noone else but us is running pods.
+			ginkgo.By("Checking the cpumanager metrics right after the kubelet restart, with pod should be admitted")
+
+			matchAlignmentMetrics := gstruct.MatchKeys(gstruct.IgnoreExtras, gstruct.Keys{
+				"kubelet_container_aligned_compute_resources_count": gstruct.MatchAllElements(nodeID, gstruct.Elements{
+					metrics.AlignedNUMANode: timelessSample(1),
+				}),
+			})
+
+			ginkgo.By("Giving the Kubelet time to update the alignment metrics")
+			gomega.Eventually(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchAlignmentMetrics)
+			ginkgo.By("Ensuring the metrics match the expectations about alignment metrics a few more times")
+			gomega.Consistently(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchAlignmentMetrics)
+		})
 	})
 })