[KNI][release-4.14] nrt: cache: add support for dedicated informer

apis/config/types.go

@@ -176,6 +176,14 @@ const (
 	CacheResyncOnlyExclusiveResources CacheResyncMethod = "OnlyExclusiveResources"
 )
 
+// CacheInformerMode is a "string" type
+type CacheInformerMode string
+
+const (
+	CacheInformerShared    CacheInformerMode = "Shared"
+	CacheInformerDedicated CacheInformerMode = "Dedicated"
+)
+
 // NodeResourceTopologyCache define configuration details for the NodeResourceTopology cache.
 type NodeResourceTopologyCache struct {
 	// ForeignPodsDetect sets how foreign pods should be handled.
@@ -192,6 +200,11 @@ type NodeResourceTopologyCache struct {
 	// Has no effect if caching is disabled (CacheResyncPeriod is zero) or if DiscardReservedNodes
 	// is enabled. "Autodetect" is the default, reads hint from NRT objects. Fallback is "All".
 	ResyncMethod *CacheResyncMethod
+	// InformerMode controls the channel the cache uses to get updates about pods.
+	// "Shared" uses the default settings; "Dedicated" creates a specific subscription which is
+	// guaranteed to best suit the cache needs, at cost of one extra connection.
+	// If unspecified, default is "Dedicated"
+	InformerMode *CacheInformerMode
 }
 
 // +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object

apis/config/v1/defaults.go

@@ -89,6 +89,8 @@ var (
 
 	defaultResyncMethod = CacheResyncAutodetect
 
+	defaultInformerMode = CacheInformerDedicated
+
 	// Defaults for NetworkOverhead
 	// DefaultWeightsName contains the default costs to be used by networkAware plugins
 	DefaultWeightsName = "UserDefined"
@@ -200,6 +202,9 @@ func SetDefaults_NodeResourceTopologyMatchArgs(obj *NodeResourceTopologyMatchArg
 	if obj.Cache.ResyncMethod == nil {
 		obj.Cache.ResyncMethod = &defaultResyncMethod
 	}
+	if obj.Cache.InformerMode == nil {
+		obj.Cache.InformerMode = &defaultInformerMode
+	}
 }
 
 // SetDefaults_PreemptionTolerationArgs reuses SetDefaults_DefaultPreemptionArgs

apis/config/v1/defaults_test.go

@@ -205,6 +205,7 @@ func TestSchedulingDefaults(t *testing.T) {
 				Cache: &NodeResourceTopologyCache{
 					ForeignPodsDetect: &defaultForeignPodsDetect,
 					ResyncMethod:      &defaultResyncMethod,
+					InformerMode:      &defaultInformerMode,
 				},
 			},
 		},

apis/config/v1/types.go

@@ -174,6 +174,14 @@ const (
 	CacheResyncOnlyExclusiveResources CacheResyncMethod = "OnlyExclusiveResources"
 )
 
+// CacheInformerMode is a "string" type
+type CacheInformerMode string
+
+const (
+	CacheInformerShared    CacheInformerMode = "Shared"
+	CacheInformerDedicated CacheInformerMode = "Dedicated"
+)
+
 // NodeResourceTopologyCache define configuration details for the NodeResourceTopology cache.
 type NodeResourceTopologyCache struct {
 	// ForeignPodsDetect sets how foreign pods should be handled.
@@ -190,6 +198,11 @@ type NodeResourceTopologyCache struct {
 	// Has no effect if caching is disabled (CacheResyncPeriod is zero) or if DiscardReservedNodes
 	// is enabled. "Autodetect" is the default, reads hint from NRT objects. Fallback is "All".
 	ResyncMethod *CacheResyncMethod `json:"resyncMethod,omitempty"`
+	// InformerMode controls the channel the cache uses to get updates about pods.
+	// "Shared" uses the default settings; "Dedicated" creates a specific subscription which is
+	// guaranteed to best suit the cache needs, at cost of one extra connection.
+	// If unspecified, default is "Dedicated"
+	InformerMode *CacheInformerMode `json:"informerMode,omitempty"`
 }
 
 // +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object

apis/config/v1beta3/defaults.go

@@ -89,6 +89,8 @@ var (
 
 	defaultResyncMethod = CacheResyncAutodetect
 
+	defaultInformerMode = CacheInformerDedicated
+
 	// Defaults for NetworkOverhead
 	// DefaultWeightsName contains the default costs to be used by networkAware plugins
 	DefaultWeightsName = "UserDefined"
@@ -200,6 +202,9 @@ func SetDefaults_NodeResourceTopologyMatchArgs(obj *NodeResourceTopologyMatchArg
 	if obj.Cache.ResyncMethod == nil {
 		obj.Cache.ResyncMethod = &defaultResyncMethod
 	}
+	if obj.Cache.InformerMode == nil {
+		obj.Cache.InformerMode = &defaultInformerMode
+	}
 }
 
 // SetDefaults_PreemptionTolerationArgs reuses SetDefaults_DefaultPreemptionArgs

apis/config/v1beta3/defaults_test.go

@@ -205,6 +205,7 @@ func TestSchedulingDefaults(t *testing.T) {
 				Cache: &NodeResourceTopologyCache{
 					ForeignPodsDetect: &defaultForeignPodsDetect,
 					ResyncMethod:      &defaultResyncMethod,
+					InformerMode:      &defaultInformerMode,
 				},
 			},
 		},

apis/config/v1beta3/types.go

@@ -174,6 +174,14 @@ const (
 	CacheResyncOnlyExclusiveResources CacheResyncMethod = "OnlyExclusiveResources"
 )
 
+// CacheInformerMode is a "string" type
+type CacheInformerMode string
+
+const (
+	CacheInformerShared    CacheInformerMode = "Shared"
+	CacheInformerDedicated CacheInformerMode = "Dedicated"
+)
+
 // NodeResourceTopologyCache define configuration details for the NodeResourceTopology cache.
 type NodeResourceTopologyCache struct {
 	// ForeignPodsDetect sets how foreign pods should be handled.
@@ -190,6 +198,11 @@ type NodeResourceTopologyCache struct {
 	// Has no effect if caching is disabled (CacheResyncPeriod is zero) or if DiscardReservedNodes
 	// is enabled. "Autodetect" is the default, reads hint from NRT objects. Fallback is "All".
 	ResyncMethod *CacheResyncMethod `json:"resyncMethod,omitempty"`
+	// InformerMode controls the channel the cache uses to get updates about pods.
+	// "Shared" uses the default settings; "Dedicated" creates a specific subscription which is
+	// guaranteed to best suit the cache needs, at cost of one extra connection.
+	// If unspecified, default is "Dedicated"
+	InformerMode *CacheInformerMode `json:"informerMode,omitempty"`
 }
 
 // +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object

cmd/noderesourcetopology-plugin/main.go

@@ -36,7 +36,6 @@ import (
 	_ "sigs.k8s.io/scheduler-plugins/apis/config/scheme"
 
 	knifeatures "sigs.k8s.io/scheduler-plugins/pkg-kni/features"
-	kniinformer "sigs.k8s.io/scheduler-plugins/pkg-kni/podinformer"
 
 	"github.com/k8stopologyawareschedwg/podfingerprint"
 )
@@ -66,8 +65,6 @@ func main() {
 
 	rand.Seed(time.Now().UnixNano())
 
-	kniinformer.Setup()
-
 	// Register custom plugins to the scheduler framework.
 	// Later they can consist of scheduler profile(s) and hence
 	// used by various kinds of workloads.

pkg/noderesourcetopology/cache/overreserve.go

@@ -27,15 +27,16 @@ import (
 	"github.com/k8stopologyawareschedwg/podfingerprint"
 
 	corev1 "k8s.io/api/core/v1"
+	"k8s.io/apimachinery/pkg/labels"
 	"k8s.io/apimachinery/pkg/types"
 	podlisterv1 "k8s.io/client-go/listers/core/v1"
-	k8scache "k8s.io/client-go/tools/cache"
 	"k8s.io/klog/v2"
-	"k8s.io/kubernetes/pkg/scheduler/framework"
 
 	ctrlclient "sigs.k8s.io/controller-runtime/pkg/client"
 
 	apiconfig "sigs.k8s.io/scheduler-plugins/apis/config"
+	"sigs.k8s.io/scheduler-plugins/pkg/noderesourcetopology/podprovider"
+	"sigs.k8s.io/scheduler-plugins/pkg/noderesourcetopology/resourcerequests"
 	"sigs.k8s.io/scheduler-plugins/pkg/noderesourcetopology/stringify"
 )
 
@@ -50,9 +51,10 @@ type OverReserve struct {
 	nodesWithForeignPods   counter
 	podLister              podlisterv1.PodLister
 	resyncMethod           apiconfig.CacheResyncMethod
+	isPodRelevant          podprovider.PodFilterFunc
 }
 
-func NewOverReserve(cfg *apiconfig.NodeResourceTopologyCache, client ctrlclient.Client, podLister podlisterv1.PodLister) (*OverReserve, error) {
+func NewOverReserve(cfg *apiconfig.NodeResourceTopologyCache, client ctrlclient.Client, podLister podlisterv1.PodLister, isPodRelevant podprovider.PodFilterFunc) (*OverReserve, error) {
 	if client == nil || podLister == nil {
 		return nil, fmt.Errorf("nrtcache: received nil references")
 	}
@@ -74,6 +76,7 @@ func NewOverReserve(cfg *apiconfig.NodeResourceTopologyCache, client ctrlclient.
 		nodesWithForeignPods:   newCounter(),
 		podLister:              podLister,
 		resyncMethod:           resyncMethod,
+		isPodRelevant:          isPodRelevant,
 	}
 	return obj, nil
 }
@@ -199,7 +202,7 @@ func (ov *OverReserve) Resync() {
 	}
 
 	// node -> pod identifier (namespace, name)
-	nodeToObjsMap, err := makeNodeToPodDataMap(ov.podLister, logID)
+	nodeToObjsMap, err := makeNodeToPodDataMap(ov.podLister, ov.isPodRelevant, logID)
 	if err != nil {
 		klog.ErrorS(err, "cannot find the mapping between running pods and nodes")
 		return
@@ -267,16 +270,32 @@ func (ov *OverReserve) FlushNodes(logID string, nrts ...*topologyv1alpha2.NodeRe
 	}
 }
 
-func InformerFromHandle(handle framework.Handle) (k8scache.SharedIndexInformer, podlisterv1.PodLister) {
-	podHandle := handle.SharedInformerFactory().Core().V1().Pods() // shortcut
-	return podHandle.Informer(), podHandle.Lister()
-}
-
 // to be used only in tests
 func (ov *OverReserve) Store() *nrtStore {
 	return ov.nrts
 }
 
+func makeNodeToPodDataMap(podLister podlisterv1.PodLister, isPodRelevant podprovider.PodFilterFunc, logID string) (map[string][]podData, error) {
+	nodeToObjsMap := make(map[string][]podData)
+	pods, err := podLister.List(labels.Everything())
+	if err != nil {
+		return nodeToObjsMap, err
+	}
+	for _, pod := range pods {
+		if !isPodRelevant(pod, logID) {
+			continue
+		}
+		nodeObjs := nodeToObjsMap[pod.Spec.NodeName]
+		nodeObjs = append(nodeObjs, podData{
+			Namespace:             pod.Namespace,
+			Name:                  pod.Name,
+			HasExclusiveResources: resourcerequests.AreExclusiveForPod(pod),
+		})
+		nodeToObjsMap[pod.Spec.NodeName] = nodeObjs
+	}
+	return nodeToObjsMap, nil
+}
+
 func logIDFromTime() string {
 	return fmt.Sprintf("resync%v", time.Now().UnixMilli())
 }