Merge pull request kubernetes#130491 from swatisehgal/cpu-allocation-numa-spread-metric

Add metrics to capture CPU allocation spread across NUMA nodes

Author: k8s-ci-robot

pkg/kubelet/cm/cpumanager/policy_static.go

@@ -18,6 +18,7 @@ package cpumanager
 
 import (
 	"fmt"
+	"strconv"
 
 	v1 "k8s.io/api/core/v1"
 	utilfeature "k8s.io/apiserver/pkg/util/feature"
@@ -389,7 +390,7 @@ func (p *staticPolicy) Allocate(s state.State, pod *v1.Pod, container *v1.Contai
 
 	s.SetCPUSet(string(pod.UID), container.Name, cpuAllocation.CPUs)
 	p.updateCPUsToReuse(pod, container, cpuAllocation.CPUs)
-	p.updateMetricsOnAllocate(cpuAllocation)
+	p.updateMetricsOnAllocate(s, cpuAllocation)
 
 	klog.V(4).InfoS("Allocated exclusive CPUs", "pod", klog.KObj(pod), "containerName", container.Name, "cpuset", cpuAllocation.CPUs.String())
 	return nil
@@ -416,7 +417,8 @@ func (p *staticPolicy) RemoveContainer(s state.State, podUID string, containerNa
 		// Mutate the shared pool, adding released cpus.
 		toRelease = toRelease.Difference(cpusInUse)
 		s.SetDefaultCPUSet(s.GetDefaultCPUSet().Union(toRelease))
-		p.updateMetricsOnRelease(toRelease)
+		p.updateMetricsOnRelease(s, toRelease)
+
 	}
 	return nil
 }
@@ -755,33 +757,60 @@ func (p *staticPolicy) getAlignedCPUs(numaAffinity bitmask.BitMask, allocatableC
 
 func (p *staticPolicy) initializeMetrics(s state.State) {
 	metrics.CPUManagerSharedPoolSizeMilliCores.Set(float64(p.GetAvailableCPUs(s).Size() * 1000))
-	metrics.CPUManagerExclusiveCPUsAllocationCount.Set(float64(countExclusiveCPUs(s)))
 	metrics.ContainerAlignedComputeResourcesFailure.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedPhysicalCPU).Add(0) // ensure the value exists
 	metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedPhysicalCPU).Add(0)        // ensure the value exists
 	metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedUncoreCache).Add(0)        // ensure the value exists
+	totalAssignedCPUs := getTotalAssignedExclusiveCPUs(s)
+	metrics.CPUManagerExclusiveCPUsAllocationCount.Set(float64(totalAssignedCPUs.Size()))
+	updateAllocationPerNUMAMetric(p.topology, totalAssignedCPUs)
 }
 
-func (p *staticPolicy) updateMetricsOnAllocate(cpuAlloc topology.Allocation) {
+func (p *staticPolicy) updateMetricsOnAllocate(s state.State, cpuAlloc topology.Allocation) {
 	ncpus := cpuAlloc.CPUs.Size()
 	metrics.CPUManagerExclusiveCPUsAllocationCount.Add(float64(ncpus))
 	metrics.CPUManagerSharedPoolSizeMilliCores.Add(float64(-ncpus * 1000))
 	if cpuAlloc.Aligned.UncoreCache {
 		metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignScopeContainer, metrics.AlignedUncoreCache).Inc()
 	}
+	totalAssignedCPUs := getTotalAssignedExclusiveCPUs(s)
+	updateAllocationPerNUMAMetric(p.topology, totalAssignedCPUs)
 }
 
-func (p *staticPolicy) updateMetricsOnRelease(cset cpuset.CPUSet) {
+func (p *staticPolicy) updateMetricsOnRelease(s state.State, cset cpuset.CPUSet) {
 	ncpus := cset.Size()
 	metrics.CPUManagerExclusiveCPUsAllocationCount.Add(float64(-ncpus))
 	metrics.CPUManagerSharedPoolSizeMilliCores.Add(float64(ncpus * 1000))
+	totalAssignedCPUs := getTotalAssignedExclusiveCPUs(s)
+	updateAllocationPerNUMAMetric(p.topology, totalAssignedCPUs.Difference(cset))
+}
+
+func getTotalAssignedExclusiveCPUs(s state.State) cpuset.CPUSet {
+	totalAssignedCPUs := cpuset.New()
+	for _, assignment := range s.GetCPUAssignments() {
+		for _, cset := range assignment {
+			totalAssignedCPUs = totalAssignedCPUs.Union(cset)
+		}
+
+	}
+	return totalAssignedCPUs
 }
 
-func countExclusiveCPUs(s state.State) int {
-	exclusiveCPUs := 0
-	for _, cpuAssign := range s.GetCPUAssignments() {
-		for _, cset := range cpuAssign {
-			exclusiveCPUs += cset.Size()
+func updateAllocationPerNUMAMetric(topo *topology.CPUTopology, allocatedCPUs cpuset.CPUSet) {
+	numaCount := make(map[int]int)
+
+	// Count CPUs allocated per NUMA node
+	for _, cpuID := range allocatedCPUs.UnsortedList() {
+		numaNode, err := topo.CPUNUMANodeID(cpuID)
+		if err != nil {
+			//NOTE: We are logging the error but it is highly unlikely to happen as the CPUset
+			//      is already computed, evaluated and there is no room for user tampering.
+			klog.ErrorS(err, "Unable to determine NUMA node", "cpuID", cpuID)
 		}
+		numaCount[numaNode]++
+	}
+
+	// Update metric
+	for numaNode, count := range numaCount {
+		metrics.CPUManagerAllocationPerNUMA.WithLabelValues(strconv.Itoa(numaNode)).Set(float64(count))
 	}
-	return exclusiveCPUs
 }

pkg/kubelet/metrics/metrics.go

@@ -113,6 +113,7 @@ const (
 	CPUManagerPinningErrorsTotalKey           = "cpu_manager_pinning_errors_total"
 	CPUManagerSharedPoolSizeMilliCoresKey     = "cpu_manager_shared_pool_size_millicores"
 	CPUManagerExclusiveCPUsAllocationCountKey = "cpu_manager_exclusive_cpu_allocation_count"
+	CPUManagerAllocationPerNUMAKey            = "cpu_manager_allocation_per_numa"
 
 	// Metrics to track the Memory manager behavior
 	MemoryManagerPinningRequestsTotalKey = "memory_manager_pinning_requests_total"
@@ -815,6 +816,17 @@ var (
 		},
 	)
 
+	// CPUManagerAllocationPerNUMA tracks the count of CPUs allocated per NUMA node
+	CPUManagerAllocationPerNUMA = metrics.NewGaugeVec(
+		&metrics.GaugeOpts{
+			Subsystem:      KubeletSubsystem,
+			Name:           CPUManagerAllocationPerNUMAKey,
+			Help:           "Number of CPUs allocated per NUMA node",
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{AlignedNUMANode},
+	)
+
 	// ContainerAlignedComputeResources reports the count of resources allocation which granted aligned resources, per alignment boundary
 	ContainerAlignedComputeResources = metrics.NewCounterVec(
 		&metrics.CounterOpts{
@@ -1126,6 +1138,7 @@ func Register(collectors ...metrics.StableCollector) {
 		legacyregistry.MustRegister(CPUManagerPinningErrorsTotal)
 		legacyregistry.MustRegister(CPUManagerSharedPoolSizeMilliCores)
 		legacyregistry.MustRegister(CPUManagerExclusiveCPUsAllocationCount)
+		legacyregistry.MustRegister(CPUManagerAllocationPerNUMA)
 		legacyregistry.MustRegister(ContainerAlignedComputeResources)
 		legacyregistry.MustRegister(ContainerAlignedComputeResourcesFailure)
 		legacyregistry.MustRegister(MemoryManagerPinningRequestTotal)

test/e2e_node/cpu_manager_metrics_test.go

@@ -34,6 +34,7 @@ import (
 	kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config"
 	"k8s.io/kubernetes/pkg/kubelet/apis/podresources"
 	"k8s.io/kubernetes/pkg/kubelet/cm/cpumanager"
+	"k8s.io/kubernetes/pkg/kubelet/metrics"
 	"k8s.io/kubernetes/pkg/kubelet/util"
 	"k8s.io/kubernetes/test/e2e/feature"
 	"k8s.io/kubernetes/test/e2e/framework"
@@ -389,6 +390,129 @@ var _ = SIGDescribe("CPU Manager Metrics", framework.WithSerial(), feature.CPUMa
 			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)
 		})
+		ginkgo.It("should report zero counters for allocation per NUMA after a fresh restart", func(ctx context.Context) {
+
+			cpuPolicyOptions := map[string]string{
+				cpumanager.DistributeCPUsAcrossNUMAOption: "true",
+				cpumanager.FullPCPUsOnlyOption:            "true",
+			}
+			newCfg := configureCPUManagerInKubelet(oldCfg,
+				&cpuManagerKubeletArguments{
+					policyName:              string(cpumanager.PolicyStatic),
+					reservedSystemCPUs:      cpuset.New(0),
+					enableCPUManagerOptions: true,
+					options:                 cpuPolicyOptions,
+				},
+			)
+
+			updateKubeletConfig(ctx, f, newCfg, true)
+
+			ginkgo.By("Checking the cpumanager allocation per NUMA metric right after the kubelet restart, with no pods running")
+			numaNodes, _, _ := hostCheck()
+
+			framework.Logf("numaNodes on the system %d", numaNodes)
+
+			keys := make(map[interface{}]types.GomegaMatcher)
+			idFn := makeCustomLabelID(metrics.AlignedNUMANode)
+
+			for i := 0; i < numaNodes; i++ {
+				keys["kubelet_cpu_manager_allocation_per_numa"] = gstruct.MatchAllElements(idFn, gstruct.Elements{
+					fmt.Sprintf("%d", i): timelessSample(0),
+				})
+
+			}
+
+			matchSpreadMetrics := gstruct.MatchKeys(gstruct.IgnoreExtras, keys)
+
+			ginkgo.By("Giving the Kubelet time to start up and produce metrics")
+			gomega.Eventually(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchSpreadMetrics)
+			ginkgo.By("Ensuring the metrics match the expectations a few more times")
+			gomega.Consistently(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchSpreadMetrics)
+
+		})
+
+		ginkgo.It("should report allocation per NUMA metric when handling guaranteed pods", func(ctx context.Context) {
+			var cpusNumPerNUMA, coresNumPerNUMA, numaNodes, threadsPerCore int
+			cpuPolicyOptions := map[string]string{
+				cpumanager.DistributeCPUsAcrossNUMAOption: "true",
+				cpumanager.FullPCPUsOnlyOption:            "true",
+			}
+			newCfg := configureCPUManagerInKubelet(oldCfg,
+				&cpuManagerKubeletArguments{
+					policyName:              string(cpumanager.PolicyStatic),
+					reservedSystemCPUs:      cpuset.New(0),
+					enableCPUManagerOptions: true,
+					options:                 cpuPolicyOptions,
+				},
+			)
+
+			updateKubeletConfig(ctx, f, newCfg, true)
+
+			numaNodes, coresNumPerNUMA, threadsPerCore = hostCheck()
+			cpusNumPerNUMA = coresNumPerNUMA * threadsPerCore
+
+			framework.Logf("numaNodes on the system %d", numaNodes)
+			framework.Logf("Cores per NUMA on the system %d", coresNumPerNUMA)
+			framework.Logf("Threads per Core on the system %d", threadsPerCore)
+			framework.Logf("CPUs per NUMA on the system %d", cpusNumPerNUMA)
+
+			smtLevel = getSMTLevel()
+			framework.Logf("SMT Level on the system %d", smtLevel)
+
+			ginkgo.By("Querying the podresources endpoint to get the baseline")
+			endpoint, err := util.LocalEndpoint(defaultPodResourcesPath, podresources.Socket)
+			framework.ExpectNoError(err, "LocalEndpoint() failed err: %v", err)
+
+			cli, conn, err := podresources.GetV1Client(endpoint, defaultPodResourcesTimeout, defaultPodResourcesMaxSize)
+			framework.ExpectNoError(err, "GetV1Client() failed err: %v", err)
+			defer func() {
+				framework.ExpectNoError(conn.Close())
+			}()
+
+			ginkgo.By("Checking the pool allocatable resources from the kubelet")
+			resp, err := cli.GetAllocatableResources(ctx, &kubeletpodresourcesv1.AllocatableResourcesRequest{})
+			framework.ExpectNoError(err, "failed to get the kubelet allocatable resources")
+			allocatableCPUs, _ := demuxCPUsAndDevicesFromGetAllocatableResources(resp)
+
+			// 'distribute-cpus-across-numa' policy option ensures that CPU allocations are evenly distributed
+			//  across NUMA nodes in cases where more than one NUMA node is required to satisfy the allocation.
+			// So, we want to ensure that the CPU Request exceeds the number of CPUs that can fit within a single
+			// NUMA node. We have to pick cpuRequest such that:
+			// 1. CPURequest > cpusNumPerNUMA
+			// 2. Not occupy all the CPUs on the node ande leave room for reserved CPU
+			// 3. CPURequest is a multiple if number of NUMA nodes to allow equal CPU distribution across NUMA nodes
+			//
+			// In summary: cpusNumPerNUMA < CPURequest < ((cpusNumPerNuma * numaNodes) - reservedCPUscount)
+			// Considering all these constraints we select: CPURequest= (cpusNumPerNUMA-smtLevel)*numaNodes
+			cpuRequest := (cpusNumPerNUMA - smtLevel) * numaNodes
+			if cpuRequest > allocatableCPUs.Size() {
+				e2eskipper.Skipf("Pod requesting %d CPUs which is more than allocatable CPUs:%d", cpuRequest, allocatableCPUs.Size())
+			}
+
+			ginkgo.By("Creating the test pod")
+			testPod = e2epod.NewPodClient(f).Create(ctx, makeGuaranteedCPUExclusiveSleeperPod("test-pod-allocation-per-numa", cpuRequest))
+
+			ginkgo.By("Checking the cpumanager metrics after pod creation")
+
+			keys := make(map[interface{}]types.GomegaMatcher)
+			idFn := makeCustomLabelID(metrics.AlignedNUMANode)
+
+			// On a clean environment with no other pods running if distribute-across-numa policy option is enabled
+			for i := 0; i < numaNodes; i++ {
+				keys["kubelet_cpu_manager_allocation_per_numa"] = gstruct.MatchAllElements(idFn, gstruct.Elements{
+					fmt.Sprintf("%d", i): timelessSample(2),
+				})
+
+			}
+
+			matchSpreadMetrics := gstruct.MatchKeys(gstruct.IgnoreExtras, keys)
+
+			ginkgo.By("Giving the Kubelet time to start up and produce metrics")
+			gomega.Eventually(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchSpreadMetrics)
+			ginkgo.By("Ensuring the metrics match the expectations a few more times")
+			gomega.Consistently(ctx, getKubeletMetrics, 1*time.Minute, 15*time.Second).Should(matchSpreadMetrics)
+		})
+
 	})
 })
 

test/e2e_node/resource_metrics_test.go

@@ -187,6 +187,13 @@ func makeCustomPairID(pri, sec string) func(interface{}) string {
 	}
 }
 
+func makeCustomLabelID(label string) func(interface{}) string {
+	return func(element interface{}) string {
+		el := element.(*model.Sample)
+		return string(el.Metric[model.LabelName(label)])
+	}
+}
+
 func boundedSample(lower, upper interface{}) types.GomegaMatcher {
 	return gstruct.PointTo(gstruct.MatchAllFields(gstruct.Fields{
 		// We already check Metric when matching the Id