experimental: l3 group metrics

wire in the l3 group alignment metrics

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

pkg/kubelet/cm/cpumanager/cpu_assignment.go

@@ -24,6 +24,7 @@ import (
 	"k8s.io/klog/v2"
 
 	"k8s.io/kubernetes/pkg/kubelet/cm/cpumanager/topology"
+	"k8s.io/kubernetes/pkg/kubelet/metrics"
 	"k8s.io/utils/cpuset"
 )
 
@@ -102,7 +103,7 @@ func min(x, y int) int {
 type numaOrSocketsFirstFuncs interface {
 	takeFullFirstLevel()
 	takeFullSecondLevel()
-	takeThirdLevel()
+	takeThirdLevel() (int, int)
 	sortAvailableNUMANodes() []int
 	sortAvailableSockets() []int
 	sortAvailableUnCoreCaches() []int
@@ -128,8 +129,8 @@ func (n *numaFirst) takeFullSecondLevel() {
 }
 
 // In Split L3 Topology, we take from the sets of uncorecache as the third level
-func (n *numaFirst) takeThirdLevel() {
-	n.acc.takeUnCoreCache()
+func (n *numaFirst) takeThirdLevel() (int, int) {
+	return n.acc.takeUnCoreCache()
 }
 
 // If NUMA nodes are higher in the memory hierarchy than sockets, then just
@@ -189,8 +190,8 @@ func (s *socketsFirst) takeFullSecondLevel() {
 	s.acc.takeFullNUMANodes()
 }
 
-func (s *socketsFirst) takeThirdLevel() {
-	s.acc.takeUnCoreCache()
+func (s *socketsFirst) takeThirdLevel() (int, int) {
+	return s.acc.takeUnCoreCache()
 }
 
 // If sockets are higher in the memory hierarchy than NUMA nodes, then we need
@@ -430,8 +431,10 @@ func (a *cpuAccumulator) takeFullUnCore() {
 
 // First try to take partial uncorecache (CCD), if available and the request size can fit w/in the uncorecache.
 // Second try to take the full CCD if available and need is at least the size of the uncorecache group.
-func (a *cpuAccumulator) takeUnCoreCache() {
+func (a *cpuAccumulator) takeUnCoreCache() (int, int) {
 	// check if SMT ON
+	part := 0
+	full := 0
 
 	for _, uncore := range a.allUnCoreCache() {
 		numCoresNeeded := a.numCPUsNeeded / a.topo.CPUsPerCore() // this is another new change
@@ -447,14 +450,18 @@ func (a *cpuAccumulator) takeUnCoreCache() {
 		if a.numCPUsNeeded == freeCPUsInUncorecache.Size() {
 			klog.V(4).InfoS("takePartialUncore: claiming cores from Uncorecache ID", "uncore", uncore)
 			a.take(freeCPUsInUncorecache)
+			part += 1
 		}
 		// take full Uncorecache if the numCPUsNeeded is greater the L3 cache size
 		a.takeFullUnCore()
+		full += 1
 
 		if a.isSatisfied() {
-			return
+			return part, full
 		}
 	}
+
+	return part, full
 }
 
 func (a *cpuAccumulator) takeFullCores() {
@@ -614,8 +621,11 @@ func takeByTopologyUnCoreCachePacked(topo *topology.CPUTopology, availableCPUs c
 
 	// 2. Acquire partial uncorecache, if there are enough CPUs available to satisfy the container requirement
 	//    Acquire the full uncorecache, if available and the container requires at least all the CPUs in the uncorecache grouping
-	acc.numaOrSocketsFirst.takeThirdLevel()
+	part, full := acc.numaOrSocketsFirst.takeThirdLevel()
 	if acc.isSatisfied() {
+		if part == 0 && full > 0 {
+			metrics.ContainerAlignedComputeResources.WithLabelValues(metrics.AlignedL3CacheGroup).Inc()
+		}
 		return acc.result, nil
 	}
 

pkg/kubelet/metrics/metrics.go

@@ -128,8 +128,9 @@ const (
 	InitContainer      = "init_container"
 	EphemeralContainer = "ephemeral_container"
 
-	AlignedPhysicalCPU = "physical_cpu"
-	AlignedNUMAZone    = "numa_zone"
+	AlignedPhysicalCPU  = "physical_cpu"
+	AlignedNUMAZone     = "numa_zone"
+	AlignedL3CacheGroup = "l3_cache_group"
 )
 
 var (
@@ -836,7 +837,7 @@ var (
 			Help:           "Cumulative number of aligned compute resources allocated to containers by alignment type.",
 			StabilityLevel: metrics.ALPHA,
 		},
-		[]string{"physical_cpu", "numa_zone"},
+		[]string{"physical_cpu", "l3_cache_group", "numa_zone"},
 	)
 )