Merge pull request #128170 from sanposhiho/async-preemption

feature(KEP-4832): asynchronous preemption

pkg/features/kube_features.go

@@ -580,6 +580,14 @@ const (
 	// which benefits to reduce the useless requeueing.
 	SchedulerQueueingHints featuregate.Feature = "SchedulerQueueingHints"
 
+	// owner: @sanposhiho
+	// kep: http://kep.k8s.io/4832
+	// alpha: v1.32
+	//
+	// Running some expensive operation within the scheduler's preemption asynchronously,
+	// which improves the scheduling latency when the preemption involves in.
+	SchedulerAsyncPreemption featuregate.Feature = "SchedulerAsyncPreemption"
+
 	// owner: @atosatto @yuanchen8911
 	// kep: http://kep.k8s.io/3902
 	//

pkg/features/versioned_kube_features.go

@@ -633,6 +633,10 @@ var defaultVersionedKubernetesFeatureGates = map[featuregate.Feature]featuregate
 		{Version: version.MustParse("1.29"), Default: false, PreRelease: featuregate.Alpha},
 	},
 
+	SchedulerAsyncPreemption: {
+		{Version: version.MustParse("1.32"), Default: false, PreRelease: featuregate.Alpha},
+	},
+
 	SchedulerQueueingHints: {
 		{Version: version.MustParse("1.28"), Default: false, PreRelease: featuregate.Beta},
 		{Version: version.MustParse("1.32"), Default: true, PreRelease: featuregate.Beta},

pkg/scheduler/apis/config/testing/defaults/defaults.go

@@ -52,6 +52,7 @@ var ExpandedPluginsV1 = &config.Plugins{
 	PreEnqueue: config.PluginSet{
 		Enabled: []config.Plugin{
 			{Name: names.SchedulingGates},
+			{Name: names.DefaultPreemption},
 		},
 	},
 	QueueSort: config.PluginSet{

pkg/scheduler/backend/queue/scheduling_queue.go

@@ -92,9 +92,11 @@ type PreEnqueueCheck func(pod *v1.Pod) bool
 type SchedulingQueue interface {
 	framework.PodNominator
 	Add(logger klog.Logger, pod *v1.Pod)
-	// Activate moves the given pods to activeQ iff they're in unschedulablePods or backoffQ.
+	// Activate moves the given pods to activeQ.
-	// The passed-in pods are originally compiled from plugins that want to activate Pods,
+	// If a pod isn't found in unschedulablePods or backoffQ and it's in-flight,
-	// by injecting the pods through a reserved CycleState struct (PodsToActivate).
+	// the wildcard event is registered so that the pod will be requeued when it comes back.
+	// But, if a pod isn't found in unschedulablePods or backoffQ and it's not in-flight (i.e., completely unknown pod),
+	// Activate would ignore the pod.
 	Activate(logger klog.Logger, pods map[string]*v1.Pod)
 	// AddUnschedulableIfNotPresent adds an unschedulable pod back to scheduling queue.
 	// The podSchedulingCycle represents the current scheduling cycle number which can be
@@ -411,9 +413,22 @@ func (p *PriorityQueue) isPodWorthRequeuing(logger klog.Logger, pInfo *framework
 	}
 
 	if event.IsWildCard() {
+		// If the wildcard event has a Pod in newObj,
+		// that indicates that the event wants to be effective for the Pod only.
+		// Specifically, EventForceActivate could have a target Pod in newObj.
+		if newObj != nil {
+			if pod, ok := newObj.(*v1.Pod); !ok || pod.UID != pInfo.Pod.UID {
+				// This wildcard event is not for this Pod.
+				if ok {
+					logger.V(6).Info("Not worth requeuing because the event is wildcard, but for another pod", "pod", klog.KObj(pInfo.Pod), "event", event.Label(), "newObj", klog.KObj(pod))
+				}
+				return queueSkip
+			}
+		}
+
 		// If the wildcard event is special one as someone wants to force all Pods to move to activeQ/backoffQ.
 		// We return queueAfterBackoff in this case, while resetting all blocked plugins.
-		logger.V(6).Info("Worth requeuing because the event is wildcard", "pod", klog.KObj(pInfo.Pod))
+		logger.V(6).Info("Worth requeuing because the event is wildcard", "pod", klog.KObj(pInfo.Pod), "event", event.Label())
 		return queueAfterBackoff
 	}
 
@@ -590,7 +605,11 @@ func (p *PriorityQueue) Add(logger klog.Logger, pod *v1.Pod) {
 	}
 }
 
-// Activate moves the given pods to activeQ iff they're in unschedulablePods or backoffQ.
+// Activate moves the given pods to activeQ.
+// If a pod isn't found in unschedulablePods or backoffQ and it's in-flight,
+// the wildcard event is registered so that the pod will be requeued when it comes back.
+// But, if a pod isn't found in unschedulablePods or backoffQ and it's not in-flight (i.e., completely unknown pod),
+// Activate would ignore the pod.
 func (p *PriorityQueue) Activate(logger klog.Logger, pods map[string]*v1.Pod) {
 	p.lock.Lock()
 	defer p.lock.Unlock()
@@ -599,7 +618,15 @@ func (p *PriorityQueue) Activate(logger klog.Logger, pods map[string]*v1.Pod) {
 	for _, pod := range pods {
 		if p.activate(logger, pod) {
 			activated = true
+			continue
 		}
+
+		// If this pod is in-flight, register the activation event (for when QHint is enabled) or update moveRequestCycle (for when QHints is disabled)
+		// so that the pod will be requeued when it comes back.
+		// Specifically in the in-tree plugins, this is for the scenario with the preemption plugin
+		// where the async preemption API calls are all done or fail at some point before the Pod comes back to the queue.
+		p.activeQ.addEventsIfPodInFlight(nil, pod, []framework.ClusterEvent{framework.EventForceActivate})
+		p.moveRequestCycle = p.activeQ.schedulingCycle()
 	}
 
 	if activated {

pkg/scheduler/backend/queue/scheduling_queue_test.go

@@ -1294,13 +1294,17 @@ func TestPriorityQueue_Delete(t *testing.T) {
 }
 
 func TestPriorityQueue_Activate(t *testing.T) {
+	metrics.Register()
 	tests := []struct {
 		name                        string
 		qPodInfoInUnschedulablePods []*framework.QueuedPodInfo
 		qPodInfoInPodBackoffQ       []*framework.QueuedPodInfo
 		qPodInActiveQ               []*v1.Pod
 		qPodInfoToActivate          *framework.QueuedPodInfo
+		qPodInInFlightPod           *v1.Pod
+		expectedInFlightEvent       *clusterEvent
 		want                        []*framework.QueuedPodInfo
+		qHintEnabled                bool
 	}{
 		{
 			name:               "pod already in activeQ",
@@ -1313,6 +1317,21 @@ func TestPriorityQueue_Activate(t *testing.T) {
 			qPodInfoToActivate: &framework.QueuedPodInfo{PodInfo: highPriNominatedPodInfo},
 			want:               []*framework.QueuedPodInfo{},
 		},
+		{
+			name:                  "[QHint] pod not in unschedulablePods/podBackoffQ but in-flight",
+			qPodInfoToActivate:    &framework.QueuedPodInfo{PodInfo: highPriNominatedPodInfo},
+			qPodInInFlightPod:     highPriNominatedPodInfo.Pod,
+			expectedInFlightEvent: &clusterEvent{oldObj: (*v1.Pod)(nil), newObj: highPriNominatedPodInfo.Pod, event: framework.EventForceActivate},
+			want:                  []*framework.QueuedPodInfo{},
+			qHintEnabled:          true,
+		},
+		{
+			name:               "[QHint] pod not in unschedulablePods/podBackoffQ and not in-flight",
+			qPodInfoToActivate: &framework.QueuedPodInfo{PodInfo: highPriNominatedPodInfo},
+			qPodInInFlightPod:  medPriorityPodInfo.Pod, // different pod is in-flight
+			want:               []*framework.QueuedPodInfo{},
+			qHintEnabled:       true,
+		},
 		{
 			name:                        "pod in unschedulablePods",
 			qPodInfoInUnschedulablePods: []*framework.QueuedPodInfo{{PodInfo: highPriNominatedPodInfo}},
@@ -1329,12 +1348,30 @@ func TestPriorityQueue_Activate(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
+			featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.SchedulerQueueingHints, tt.qHintEnabled)
 			var objs []runtime.Object
 			logger, ctx := ktesting.NewTestContext(t)
 			ctx, cancel := context.WithCancel(ctx)
 			defer cancel()
 			q := NewTestQueueWithObjects(ctx, newDefaultQueueSort(), objs)
 
+			if tt.qPodInInFlightPod != nil {
+				// Put -> Pop the Pod to make it registered in inFlightPods.
+				q.activeQ.underLock(func(unlockedActiveQ unlockedActiveQueuer) {
+					unlockedActiveQ.AddOrUpdate(newQueuedPodInfoForLookup(tt.qPodInInFlightPod))
+				})
+				p, err := q.activeQ.pop(logger)
+				if err != nil {
+					t.Fatalf("Pop failed: %v", err)
+				}
+				if p.Pod.Name != tt.qPodInInFlightPod.Name {
+					t.Errorf("Unexpected popped pod: %v", p.Pod.Name)
+				}
+				if len(q.activeQ.listInFlightEvents()) != 1 {
+					t.Fatal("Expected the pod to be recorded in in-flight events, but it doesn't")
+				}
+			}
+
 			// Prepare activeQ/unschedulablePods/podBackoffQ according to the table
 			for _, qPod := range tt.qPodInActiveQ {
 				q.Add(logger, qPod)
@@ -1353,7 +1390,29 @@ func TestPriorityQueue_Activate(t *testing.T) {
 
 			// Check the result after activation by the length of activeQ
 			if wantLen := len(tt.want); q.activeQ.len() != wantLen {
-				t.Errorf("length compare: want %v, got %v", wantLen, q.activeQ.len())
+				t.Fatalf("length compare: want %v, got %v", wantLen, q.activeQ.len())
+			}
+
+			if tt.expectedInFlightEvent != nil {
+				if len(q.activeQ.listInFlightEvents()) != 2 {
+					t.Fatalf("Expected two in-flight event to be recorded, but got %v events", len(q.activeQ.listInFlightEvents()))
+				}
+				found := false
+				for _, e := range q.activeQ.listInFlightEvents() {
+					event, ok := e.(*clusterEvent)
+					if !ok {
+						continue
+					}
+
+					if d := cmp.Diff(tt.expectedInFlightEvent, event, cmpopts.EquateComparable(clusterEvent{})); d != "" {
+						t.Fatalf("Unexpected in-flight event (-want, +got):\n%s", d)
+					}
+					found = true
+				}
+
+				if !found {
+					t.Fatalf("Expected in-flight event to be recorded, but it wasn't.")
+				}
 			}
 
 			// Check if the specific pod exists in activeQ
@@ -3779,6 +3838,7 @@ func mustNewPodInfo(pod *v1.Pod) *framework.PodInfo {
 
 // Test_isPodWorthRequeuing tests isPodWorthRequeuing function.
 func Test_isPodWorthRequeuing(t *testing.T) {
+	metrics.Register()
 	count := 0
 	queueHintReturnQueue := func(logger klog.Logger, pod *v1.Pod, oldObj, newObj interface{}) (framework.QueueingHint, error) {
 		count++
@@ -3857,11 +3917,37 @@ func Test_isPodWorthRequeuing(t *testing.T) {
 			},
 			event:                  framework.EventUnschedulableTimeout,
 			oldObj:                 nil,
-			newObj:                 st.MakeNode().Obj(),
+			newObj:                 nil,
 			expected:               queueAfterBackoff,
 			expectedExecutionCount: 0,
 			queueingHintMap:        QueueingHintMapPerProfile{},
 		},
+		{
+			name: "return Queue when the event is wildcard and the wildcard targets the pod to be requeued right now",
+			podInfo: &framework.QueuedPodInfo{
+				UnschedulablePlugins: sets.New("fooPlugin1"),
+				PodInfo:              mustNewPodInfo(st.MakePod().Name("pod1").Namespace("ns1").UID("1").Obj()),
+			},
+			event:                  framework.EventForceActivate,
+			oldObj:                 nil,
+			newObj:                 st.MakePod().Name("pod1").Namespace("ns1").UID("1").Obj(),
+			expected:               queueAfterBackoff,
+			expectedExecutionCount: 0,
+			queueingHintMap:        QueueingHintMapPerProfile{},
+		},
+		{
+			name: "return Skip when the event is wildcard, but the wildcard targets a different pod",
+			podInfo: &framework.QueuedPodInfo{
+				UnschedulablePlugins: sets.New("fooPlugin1"),
+				PodInfo:              mustNewPodInfo(st.MakePod().Name("pod1").Namespace("ns1").UID("1").Obj()),
+			},
+			event:                  framework.EventForceActivate,
+			oldObj:                 nil,
+			newObj:                 st.MakePod().Name("pod-different").Namespace("ns2").UID("2").Obj(),
+			expected:               queueSkip,
+			expectedExecutionCount: 0,
+			queueingHintMap:        QueueingHintMapPerProfile{},
+		},
 		{
 			name: "interprets Queue from the Pending plugin as queueImmediately",
 			podInfo: &framework.QueuedPodInfo{

pkg/scheduler/framework/events.go

@@ -34,6 +34,9 @@ const (
 	// ForceActivate is the event when a pod is moved from unschedulablePods/backoffQ
 	// to activeQ. Usually it's triggered by plugin implementations.
 	ForceActivate = "ForceActivate"
+	// UnschedulableTimeout is the event when a pod is moved from unschedulablePods
+	// due to the timeout specified at pod-max-in-unschedulable-pods-duration.
+	UnschedulableTimeout = "UnschedulableTimeout"
 )
 
 var (
@@ -50,7 +53,9 @@ var (
 	// EventUnscheduledPodDelete is the event when an unscheduled pod is deleted.
 	EventUnscheduledPodDelete = ClusterEvent{Resource: unschedulablePod, ActionType: Delete}
 	// EventUnschedulableTimeout is the event when a pod stays in unschedulable for longer than timeout.
-	EventUnschedulableTimeout = ClusterEvent{Resource: WildCard, ActionType: All, label: "UnschedulableTimeout"}
+	EventUnschedulableTimeout = ClusterEvent{Resource: WildCard, ActionType: All, label: UnschedulableTimeout}
+	// EventForceActivate is the event when a pod is moved from unschedulablePods/backoffQ to activeQ.
+	EventForceActivate = ClusterEvent{Resource: WildCard, ActionType: All, label: ForceActivate}
 )
 
 // PodSchedulingPropertiesChange interprets the update of a pod and returns corresponding UpdatePodXYZ event(s).

pkg/scheduler/framework/interface.go

@@ -770,6 +770,8 @@ type Framework interface {
 
 	// SetPodNominator sets the PodNominator
 	SetPodNominator(nominator PodNominator)
+	// SetPodActivator sets the PodActivator
+	SetPodActivator(activator PodActivator)
 
 	// Close calls Close method of each plugin.
 	Close() error
@@ -783,6 +785,8 @@ type Handle interface {
 	PodNominator
 	// PluginsRunner abstracts operations to run some plugins.
 	PluginsRunner
+	// PodActivator abstracts operations in the scheduling queue.
+	PodActivator
 	// SnapshotSharedLister returns listers from the latest NodeInfo Snapshot. The snapshot
 	// is taken at the beginning of a scheduling cycle and remains unchanged until
 	// a pod finishes "Permit" point.
@@ -896,6 +900,16 @@ func (ni *NominatingInfo) Mode() NominatingMode {
 	return ni.NominatingMode
 }
 
+// PodActivator abstracts operations in the scheduling queue.
+type PodActivator interface {
+	// Activate moves the given pods to activeQ.
+	// If a pod isn't found in unschedulablePods or backoffQ and it's in-flight,
+	// the wildcard event is registered so that the pod will be requeued when it comes back.
+	// But, if a pod isn't found in unschedulablePods or backoffQ and it's not in-flight (i.e., completely unknown pod),
+	// Activate would ignore the pod.
+	Activate(logger klog.Logger, pods map[string]*v1.Pod)
+}
+
 // PodNominator abstracts operations to maintain nominated Pods.
 type PodNominator interface {
 	// AddNominatedPod adds the given pod to the nominator or

pkg/scheduler/framework/plugins/defaultpreemption/default_preemption.go

@@ -53,9 +53,11 @@ type DefaultPreemption struct {
 	args      config.DefaultPreemptionArgs
 	podLister corelisters.PodLister
 	pdbLister policylisters.PodDisruptionBudgetLister
+	Evaluator *preemption.Evaluator
 }
 
 var _ framework.PostFilterPlugin = &DefaultPreemption{}
+var _ framework.PreEnqueuePlugin = &DefaultPreemption{}
 
 // Name returns name of the plugin. It is used in logs, etc.
 func (pl *DefaultPreemption) Name() string {
@@ -71,13 +73,19 @@ func New(_ context.Context, dpArgs runtime.Object, fh framework.Handle, fts feat
 	if err := validation.ValidateDefaultPreemptionArgs(nil, args); err != nil {
 		return nil, err
 	}
+
+	podLister := fh.SharedInformerFactory().Core().V1().Pods().Lister()
+	pdbLister := getPDBLister(fh.SharedInformerFactory())
+
 	pl := DefaultPreemption{
 		fh:        fh,
 		fts:       fts,
 		args:      *args,
-		podLister: fh.SharedInformerFactory().Core().V1().Pods().Lister(),
+		podLister: podLister,
-		pdbLister: getPDBLister(fh.SharedInformerFactory()),
+		pdbLister: pdbLister,
 	}
+	pl.Evaluator = preemption.NewEvaluator(Name, fh, &pl, fts.EnableAsyncPreemption)
+
 	return &pl, nil
 }
 
@@ -87,16 +95,7 @@ func (pl *DefaultPreemption) PostFilter(ctx context.Context, state *framework.Cy
 		metrics.PreemptionAttempts.Inc()
 	}()
 
-	pe := preemption.Evaluator{
+	result, status := pl.Evaluator.Preempt(ctx, state, pod, m)
-		PluginName: names.DefaultPreemption,
-		Handler:    pl.fh,
-		PodLister:  pl.podLister,
-		PdbLister:  pl.pdbLister,
-		State:      state,
-		Interface:  pl,
-	}
-
-	result, status := pe.Preempt(ctx, pod, m)
 	msg := status.Message()
 	if len(msg) > 0 {
 		return result, framework.NewStatus(status.Code(), "preemption: "+msg)
@@ -104,6 +103,24 @@ func (pl *DefaultPreemption) PostFilter(ctx context.Context, state *framework.Cy
 	return result, status
 }
 
+func (pl *DefaultPreemption) PreEnqueue(ctx context.Context, p *v1.Pod) *framework.Status {
+	if !pl.fts.EnableAsyncPreemption {
+		return nil
+	}
+	if pl.Evaluator.IsPodRunningPreemption(p.GetUID()) {
+		return framework.NewStatus(framework.UnschedulableAndUnresolvable, "waiting for the preemption for this pod to be finished")
+	}
+	return nil
+}
+
+// EventsToRegister returns the possible events that may make a Pod
+// failed by this plugin schedulable.
+func (pl *DefaultPreemption) EventsToRegister(_ context.Context) ([]framework.ClusterEventWithHint, error) {
+	// The plugin moves the preemptor Pod to acviteQ/backoffQ once the preemption API calls are all done,
+	// and we don't need to move the Pod with any events.
+	return nil, nil
+}
+
 // calculateNumCandidates returns the number of candidates the FindCandidates
 // method must produce from dry running based on the constraints given by
 // <minCandidateNodesPercentage> and <minCandidateNodesAbsolute>. The number of

pkg/scheduler/framework/plugins/defaultpreemption/default_preemption_test.go

@@ -25,6 +25,7 @@ import (
 	"math/rand"
 	"sort"
 	"strings"
+	"sync"
 	"testing"
 	"time"
 
@@ -37,10 +38,12 @@ import (
 	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/util/sets"
 	"k8s.io/apimachinery/pkg/util/strategicpatch"
+	"k8s.io/apimachinery/pkg/util/wait"
 	"k8s.io/client-go/informers"
 	clientsetfake "k8s.io/client-go/kubernetes/fake"
 	clienttesting "k8s.io/client-go/testing"
 	"k8s.io/client-go/tools/events"
+	"k8s.io/klog/v2"
 	"k8s.io/klog/v2/ktesting"
 	kubeschedulerconfigv1 "k8s.io/kube-scheduler/config/v1"
 	extenderv1 "k8s.io/kube-scheduler/extender/v1"
@@ -436,6 +439,7 @@ func TestPostFilter(t *testing.T) {
 				pdbLister: getPDBLister(informerFactory),
 				args:      *getDefaultDefaultPreemptionArgs(),
 			}
+			p.Evaluator = preemption.NewEvaluator(names.DefaultPreemption, f, &p, false)
 
 			state := framework.NewCycleState()
 			// Ensure <state> is populated.
@@ -1206,11 +1210,10 @@ func TestDryRunPreemption(t *testing.T) {
 					Handler:    pl.fh,
 					PodLister:  pl.podLister,
 					PdbLister:  pl.pdbLister,
-					State:      state,
 					Interface:  pl,
 				}
 				offset, numCandidates := pl.GetOffsetAndNumCandidates(int32(len(nodeInfos)))
-				got, _, _ := pe.DryRunPreemption(ctx, pod, nodeInfos, tt.pdbs, offset, numCandidates)
+				got, _, _ := pe.DryRunPreemption(ctx, state, pod, nodeInfos, tt.pdbs, offset, numCandidates)
 				// Sort the values (inner victims) and the candidate itself (by its NominatedNodeName).
 				for i := range got {
 					victims := got[i].Victims().Pods
@@ -1447,11 +1450,10 @@ func TestSelectBestCandidate(t *testing.T) {
 				Handler:    pl.fh,
 				PodLister:  pl.podLister,
 				PdbLister:  pl.pdbLister,
-				State:      state,
 				Interface:  pl,
 			}
 			offset, numCandidates := pl.GetOffsetAndNumCandidates(int32(len(nodeInfos)))
-			candidates, _, _ := pe.DryRunPreemption(ctx, tt.pod, nodeInfos, nil, offset, numCandidates)
+			candidates, _, _ := pe.DryRunPreemption(ctx, state, tt.pod, nodeInfos, nil, offset, numCandidates)
 			s := pe.SelectCandidate(ctx, candidates)
 			if s == nil || len(s.Name()) == 0 {
 				return
@@ -1548,7 +1550,9 @@ func TestPodEligibleToPreemptOthers(t *testing.T) {
 		})
 	}
 }
+
 func TestPreempt(t *testing.T) {
+	metrics.Register()
 	tests := []struct {
 		name           string
 		pod            *v1.Pod
@@ -1713,16 +1717,29 @@ func TestPreempt(t *testing.T) {
 	}
 
 	labelKeys := []string{"hostname", "zone", "region"}
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				client := clientsetfake.NewClientset()
 				informerFactory := informers.NewSharedInformerFactory(client, 0)
 				podInformer := informerFactory.Core().V1().Pods().Informer()
-			podInformer.GetStore().Add(test.pod)
+				testPod := test.pod.DeepCopy()
+				testPods := make([]*v1.Pod, len(test.pods))
 				for i := range test.pods {
-				podInformer.GetStore().Add(test.pods[i])
+					testPods[i] = test.pods[i].DeepCopy()
 				}
 
+				if err := podInformer.GetStore().Add(testPod); err != nil {
+					t.Fatalf("Failed to add test pod %s: %v", testPod.Name, err)
+				}
+				for i := range testPods {
+					if err := podInformer.GetStore().Add(testPods[i]); err != nil {
+						t.Fatalf("Failed to add test pod %s: %v", testPods[i], err)
+					}
+				}
+
+				// Need to protect deletedPodNames and patchedPodNames to prevent DATA RACE panic.
+				var mu sync.RWMutex
 				deletedPodNames := sets.New[string]()
 				patchedPodNames := sets.New[string]()
 				patchedPods := []*v1.Pod{}
@@ -1748,10 +1765,14 @@ func TestPreempt(t *testing.T) {
 						t.Fatalf("Failed to unmarshal updated pod %q: %v", updated, err)
 					}
 					patchedPods = append(patchedPods, updatedPod)
+					mu.Lock()
+					defer mu.Unlock()
 					patchedPodNames.Insert(podName)
 					return true, nil, nil
 				})
 				client.PrependReactor("delete", "pods", func(action clienttesting.Action) (bool, runtime.Object, error) {
+					mu.Lock()
+					defer mu.Unlock()
 					deletedPodNames.Insert(action.(clienttesting.DeleteAction).GetName())
 					return true, nil, nil
 				})
@@ -1763,8 +1784,10 @@ func TestPreempt(t *testing.T) {
 				waitingPods := frameworkruntime.NewWaitingPodsMap()
 
 				cache := internalcache.New(ctx, time.Duration(0))
-			for _, pod := range test.pods {
+				for _, pod := range testPods {
-				cache.AddPod(logger, pod)
+					if err := cache.AddPod(logger, pod.DeepCopy()); err != nil {
+						t.Fatalf("Failed to add pod %s: %v", pod.Name, err)
+					}
 				}
 				cachedNodeInfoMap := map[string]*framework.NodeInfo{}
 				nodes := make([]*v1.Node, len(test.nodeNames))
@@ -1777,6 +1800,7 @@ func TestPreempt(t *testing.T) {
 						node.ObjectMeta.Labels[labelKeys[i]] = label
 					}
 					node.Name = node.ObjectMeta.Labels["hostname"]
+					t.Logf("node is added: %v. labels: %#v", node.Name, node.ObjectMeta.Labels)
 					cache.AddNode(logger, node)
 					nodes[i] = node
 
@@ -1803,10 +1827,11 @@ func TestPreempt(t *testing.T) {
 					frameworkruntime.WithEventRecorder(&events.FakeRecorder{}),
 					frameworkruntime.WithExtenders(extenders),
 					frameworkruntime.WithPodNominator(internalqueue.NewSchedulingQueue(nil, informerFactory)),
-				frameworkruntime.WithSnapshotSharedLister(internalcache.NewSnapshot(test.pods, nodes)),
+					frameworkruntime.WithSnapshotSharedLister(internalcache.NewSnapshot(testPods, nodes)),
 					frameworkruntime.WithInformerFactory(informerFactory),
 					frameworkruntime.WithWaitingPods(waitingPods),
 					frameworkruntime.WithLogger(logger),
+					frameworkruntime.WithPodActivator(&fakePodActivator{}),
 				)
 				if err != nil {
 					t.Fatal(err)
@@ -1814,7 +1839,7 @@ func TestPreempt(t *testing.T) {
 
 				state := framework.NewCycleState()
 				// Some tests rely on PreFilter plugin to compute its CycleState.
-			if _, s, _ := fwk.RunPreFilterPlugins(ctx, state, test.pod); !s.IsSuccess() {
+				if _, s, _ := fwk.RunPreFilterPlugins(ctx, state, testPod); !s.IsSuccess() {
 					t.Errorf("Unexpected preFilterStatus: %v", s)
 				}
 				// Call preempt and check the expected results.
@@ -1825,14 +1850,7 @@ func TestPreempt(t *testing.T) {
 					args:      *getDefaultDefaultPreemptionArgs(),
 				}
 
-			pe := preemption.Evaluator{
+				pe := preemption.NewEvaluator(names.DefaultPreemption, pl.fh, &pl, asyncPreemptionEnabled)
-				PluginName: names.DefaultPreemption,
-				Handler:    pl.fh,
-				PodLister:  pl.podLister,
-				PdbLister:  pl.pdbLister,
-				State:      state,
-				Interface:  &pl,
-			}
 
 				// so that these nodes are eligible for preemption, we set their status
 				// to Unschedulable.
@@ -1842,16 +1860,33 @@ func TestPreempt(t *testing.T) {
 					nodeToStatusMap.Set(n.Name, framework.NewStatus(framework.Unschedulable))
 				}
 
-			res, status := pe.Preempt(ctx, test.pod, nodeToStatusMap)
+				res, status := pe.Preempt(ctx, state, testPod, nodeToStatusMap)
 				if !status.IsSuccess() && !status.IsRejected() {
 					t.Errorf("unexpected error in preemption: %v", status.AsError())
 				}
 				if diff := cmp.Diff(test.want, res); diff != "" {
 					t.Errorf("Unexpected status (-want, +got):\n%s", diff)
 				}
-			if len(deletedPodNames) != len(test.expectedPods) {
+
-				t.Errorf("expected %v pods, got %v.", len(test.expectedPods), len(deletedPodNames))
+				if asyncPreemptionEnabled {
+					// Wait for the pod to be deleted.
+					if err := wait.PollUntilContextTimeout(ctx, time.Millisecond*200, wait.ForeverTestTimeout, false, func(ctx context.Context) (bool, error) {
+						mu.RLock()
+						defer mu.RUnlock()
+						return len(deletedPodNames) == len(test.expectedPods), nil
+					}); err != nil {
+						t.Errorf("expected %v pods to be deleted, got %v.", len(test.expectedPods), len(deletedPodNames))
 					}
+				} else {
+					mu.RLock()
+					// If async preemption is disabled, the pod should be deleted immediately.
+					if len(deletedPodNames) != len(test.expectedPods) {
+						t.Errorf("expected %v pods to be deleted, got %v.", len(test.expectedPods), len(deletedPodNames))
+					}
+					mu.RUnlock()
+				}
+
+				mu.RLock()
 				if diff := cmp.Diff(sets.List(patchedPodNames), sets.List(deletedPodNames)); diff != "" {
 					t.Errorf("unexpected difference in the set of patched and deleted pods: %s", diff)
 				}
@@ -1884,20 +1919,21 @@ func TestPreempt(t *testing.T) {
 					}
 				}
 				if res != nil && res.NominatingInfo != nil {
-				test.pod.Status.NominatedNodeName = res.NominatedNodeName
+					testPod.Status.NominatedNodeName = res.NominatedNodeName
 				}
 
 				// Manually set the deleted Pods' deletionTimestamp to non-nil.
-			for _, pod := range test.pods {
+				for _, pod := range testPods {
 					if deletedPodNames.Has(pod.Name) {
 						now := metav1.Now()
 						pod.DeletionTimestamp = &now
 						deletedPodNames.Delete(pod.Name)
 					}
 				}
+				mu.RUnlock()
 
 				// Call preempt again and make sure it doesn't preempt any more pods.
-			res, status = pe.Preempt(ctx, test.pod, framework.NewDefaultNodeToStatus())
+				res, status = pe.Preempt(ctx, state, testPod, framework.NewDefaultNodeToStatus())
 				if !status.IsSuccess() && !status.IsRejected() {
 					t.Errorf("unexpected error in preemption: %v", status.AsError())
 				}
@@ -1907,3 +1943,9 @@ func TestPreempt(t *testing.T) {
 			})
 		}
 	}
+}
+
+type fakePodActivator struct {
+}
+
+func (f *fakePodActivator) Activate(logger klog.Logger, pods map[string]*v1.Pod) {}

pkg/scheduler/framework/plugins/feature/feature.go

@@ -28,4 +28,5 @@ type Features struct {
 	EnableInPlacePodVerticalScaling              bool
 	EnableSidecarContainers                      bool
 	EnableSchedulingQueueHint                    bool
+	EnableAsyncPreemption                        bool
 }

pkg/scheduler/framework/plugins/registry.go

@@ -54,6 +54,7 @@ func NewInTreeRegistry() runtime.Registry {
 		EnableInPlacePodVerticalScaling:              feature.DefaultFeatureGate.Enabled(features.InPlacePodVerticalScaling),
 		EnableSidecarContainers:                      feature.DefaultFeatureGate.Enabled(features.SidecarContainers),
 		EnableSchedulingQueueHint:                    feature.DefaultFeatureGate.Enabled(features.SchedulerQueueingHints),
+		EnableAsyncPreemption:                        feature.DefaultFeatureGate.Enabled(features.SchedulerAsyncPreemption),
 	}
 
 	registry := runtime.Registry{

pkg/scheduler/framework/preemption/preemption.go

@@ -23,11 +23,15 @@ import (
 	"math"
 	"sync"
 	"sync/atomic"
+	"time"
 
 	v1 "k8s.io/api/core/v1"
 	policy "k8s.io/api/policy/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
 	"k8s.io/apimachinery/pkg/labels"
+	"k8s.io/apimachinery/pkg/types"
 	utilerrors "k8s.io/apimachinery/pkg/util/errors"
+	"k8s.io/apimachinery/pkg/util/sets"
 	corelisters "k8s.io/client-go/listers/core/v1"
 	policylisters "k8s.io/client-go/listers/policy/v1"
 	corev1helpers "k8s.io/component-helpers/scheduling/corev1"
@@ -36,6 +40,7 @@ import (
 	apipod "k8s.io/kubernetes/pkg/api/v1/pod"
 	"k8s.io/kubernetes/pkg/scheduler/framework"
 	"k8s.io/kubernetes/pkg/scheduler/framework/parallelize"
+	"k8s.io/kubernetes/pkg/scheduler/framework/plugins/names"
 	"k8s.io/kubernetes/pkg/scheduler/metrics"
 	"k8s.io/kubernetes/pkg/scheduler/util"
 )
@@ -125,10 +130,88 @@ type Evaluator struct {
 	Handler    framework.Handle
 	PodLister  corelisters.PodLister
 	PdbLister  policylisters.PodDisruptionBudgetLister
-	State      *framework.CycleState
+
+	enableAsyncPreemption bool
+	mu                    sync.RWMutex
+	// preempting is a set that records the pods that are currently triggering preemption asynchronously,
+	// which is used to prevent the pods from entering the scheduling cycle meanwhile.
+	preempting sets.Set[types.UID]
+
+	// PreemptPod is a function that actually makes API calls to preempt a specific Pod.
+	// This is exposed to be replaced during tests.
+	PreemptPod func(ctx context.Context, c Candidate, preemptor, victim *v1.Pod, pluginName string) error
+
 	Interface
 }
 
+func NewEvaluator(pluginName string, fh framework.Handle, i Interface, enableAsyncPreemption bool) *Evaluator {
+	podLister := fh.SharedInformerFactory().Core().V1().Pods().Lister()
+	pdbLister := fh.SharedInformerFactory().Policy().V1().PodDisruptionBudgets().Lister()
+
+	ev := &Evaluator{
+		PluginName:            names.DefaultPreemption,
+		Handler:               fh,
+		PodLister:             podLister,
+		PdbLister:             pdbLister,
+		Interface:             i,
+		enableAsyncPreemption: enableAsyncPreemption,
+		preempting:            sets.New[types.UID](),
+	}
+
+	// PreemptPod actually makes API calls to preempt a specific Pod.
+	//
+	// We implement it here directly, rather than creating a separate method like ev.preemptPod(...)
+	// to prevent the misuse of the PreemptPod function.
+	ev.PreemptPod = func(ctx context.Context, c Candidate, preemptor, victim *v1.Pod, pluginName string) error {
+		logger := klog.FromContext(ctx)
+
+		// If the victim is a WaitingPod, send a reject message to the PermitPlugin.
+		// Otherwise we should delete the victim.
+		if waitingPod := ev.Handler.GetWaitingPod(victim.UID); waitingPod != nil {
+			waitingPod.Reject(pluginName, "preempted")
+			logger.V(2).Info("Preemptor pod rejected a waiting pod", "preemptor", klog.KObj(preemptor), "waitingPod", klog.KObj(victim), "node", c.Name())
+		} else {
+			condition := &v1.PodCondition{
+				Type:    v1.DisruptionTarget,
+				Status:  v1.ConditionTrue,
+				Reason:  v1.PodReasonPreemptionByScheduler,
+				Message: fmt.Sprintf("%s: preempting to accommodate a higher priority pod", preemptor.Spec.SchedulerName),
+			}
+			newStatus := victim.Status.DeepCopy()
+			updated := apipod.UpdatePodCondition(newStatus, condition)
+			if updated {
+				if err := util.PatchPodStatus(ctx, ev.Handler.ClientSet(), victim, newStatus); err != nil {
+					logger.Error(err, "Could not add DisruptionTarget condition due to preemption", "pod", klog.KObj(victim), "preemptor", klog.KObj(preemptor))
+					return err
+				}
+			}
+			if err := util.DeletePod(ctx, ev.Handler.ClientSet(), victim); err != nil {
+				if !apierrors.IsNotFound(err) {
+					logger.Error(err, "Tried to preempted pod", "pod", klog.KObj(victim), "preemptor", klog.KObj(preemptor))
+					return err
+				}
+				logger.V(2).Info("Victim Pod is already deleted", "preemptor", klog.KObj(preemptor), "victim", klog.KObj(victim), "node", c.Name())
+				return nil
+			}
+			logger.V(2).Info("Preemptor Pod preempted victim Pod", "preemptor", klog.KObj(preemptor), "victim", klog.KObj(victim), "node", c.Name())
+		}
+
+		ev.Handler.EventRecorder().Eventf(victim, preemptor, v1.EventTypeNormal, "Preempted", "Preempting", "Preempted by pod %v on node %v", preemptor.UID, c.Name())
+
+		return nil
+	}
+
+	return ev
+}
+
+// IsPodRunningPreemption returns true if the pod is currently triggering preemption asynchronously.
+func (ev *Evaluator) IsPodRunningPreemption(podUID types.UID) bool {
+	ev.mu.RLock()
+	defer ev.mu.RUnlock()
+
+	return ev.preempting.Has(podUID)
+}
+
 // Preempt returns a PostFilterResult carrying suggested nominatedNodeName, along with a Status.
 // The semantics of returned <PostFilterResult, Status> varies on different scenarios:
 //
@@ -145,7 +228,7 @@ type Evaluator struct {
 //
 //   - <non-nil PostFilterResult, Success>. It's the regular happy path
 //     and the non-empty nominatedNodeName will be applied to the preemptor pod.
-func (ev *Evaluator) Preempt(ctx context.Context, pod *v1.Pod, m framework.NodeToStatusReader) (*framework.PostFilterResult, *framework.Status) {
+func (ev *Evaluator) Preempt(ctx context.Context, state *framework.CycleState, pod *v1.Pod, m framework.NodeToStatusReader) (*framework.PostFilterResult, *framework.Status) {
 	logger := klog.FromContext(ctx)
 
 	// 0) Fetch the latest version of <pod>.
@@ -171,7 +254,7 @@ func (ev *Evaluator) Preempt(ctx context.Context, pod *v1.Pod, m framework.NodeT
 	if err != nil {
 		return nil, framework.AsStatus(err)
 	}
-	candidates, nodeToStatusMap, err := ev.findCandidates(ctx, allNodes, pod, m)
+	candidates, nodeToStatusMap, err := ev.findCandidates(ctx, state, allNodes, pod, m)
 	if err != nil && len(candidates) == 0 {
 		return nil, framework.AsStatus(err)
 	}
@@ -203,17 +286,23 @@ func (ev *Evaluator) Preempt(ctx context.Context, pod *v1.Pod, m framework.NodeT
 		return nil, framework.NewStatus(framework.Unschedulable, "no candidate node for preemption")
 	}
 
+	logger.V(2).Info("the target node for the preemption is determined", "node", bestCandidate.Name(), "pod", klog.KObj(pod))
+
 	// 5) Perform preparation work before nominating the selected candidate.
+	if ev.enableAsyncPreemption {
+		ev.prepareCandidateAsync(bestCandidate, pod, ev.PluginName)
+	} else {
 		if status := ev.prepareCandidate(ctx, bestCandidate, pod, ev.PluginName); !status.IsSuccess() {
 			return nil, status
 		}
+	}
 
 	return framework.NewPostFilterResultWithNominatedNode(bestCandidate.Name()), framework.NewStatus(framework.Success)
 }
 
 // FindCandidates calculates a slice of preemption candidates.
 // Each candidate is executable to make the given <pod> schedulable.
-func (ev *Evaluator) findCandidates(ctx context.Context, allNodes []*framework.NodeInfo, pod *v1.Pod, m framework.NodeToStatusReader) ([]Candidate, *framework.NodeToStatus, error) {
+func (ev *Evaluator) findCandidates(ctx context.Context, state *framework.CycleState, allNodes []*framework.NodeInfo, pod *v1.Pod, m framework.NodeToStatusReader) ([]Candidate, *framework.NodeToStatus, error) {
 	if len(allNodes) == 0 {
 		return nil, nil, errors.New("no nodes available")
 	}
@@ -239,7 +328,7 @@ func (ev *Evaluator) findCandidates(ctx context.Context, allNodes []*framework.N
 	}
 
 	offset, candidatesNum := ev.GetOffsetAndNumCandidates(int32(len(potentialNodes)))
-	return ev.DryRunPreemption(ctx, pod, potentialNodes, pdbs, offset, candidatesNum)
+	return ev.DryRunPreemption(ctx, state, pod, potentialNodes, pdbs, offset, candidatesNum)
 }
 
 // callExtenders calls given <extenders> to select the list of feasible candidates.
@@ -347,41 +436,11 @@ func (ev *Evaluator) prepareCandidate(ctx context.Context, c Candidate, pod *v1.
 	defer cancel()
 	logger := klog.FromContext(ctx)
 	errCh := parallelize.NewErrorChannel()
-	preemptPod := func(index int) {
+	fh.Parallelizer().Until(ctx, len(c.Victims().Pods), func(index int) {
-		victim := c.Victims().Pods[index]
+		if err := ev.PreemptPod(ctx, c, pod, c.Victims().Pods[index], pluginName); err != nil {
-		// If the victim is a WaitingPod, send a reject message to the PermitPlugin.
-		// Otherwise we should delete the victim.
-		if waitingPod := fh.GetWaitingPod(victim.UID); waitingPod != nil {
-			waitingPod.Reject(pluginName, "preempted")
-			logger.V(2).Info("Preemptor pod rejected a waiting pod", "preemptor", klog.KObj(pod), "waitingPod", klog.KObj(victim), "node", c.Name())
-		} else {
-			condition := &v1.PodCondition{
-				Type:    v1.DisruptionTarget,
-				Status:  v1.ConditionTrue,
-				Reason:  v1.PodReasonPreemptionByScheduler,
-				Message: fmt.Sprintf("%s: preempting to accommodate a higher priority pod", pod.Spec.SchedulerName),
-			}
-			newStatus := victim.Status.DeepCopy()
-			updated := apipod.UpdatePodCondition(newStatus, condition)
-			if updated {
-				if err := util.PatchPodStatus(ctx, cs, victim, newStatus); err != nil {
-					logger.Error(err, "Could not add DisruptionTarget condition due to preemption", "pod", klog.KObj(victim), "preemptor", klog.KObj(pod))
 			errCh.SendErrorWithCancel(err, cancel)
-					return
 		}
-			}
+	}, ev.PluginName)
-			if err := util.DeletePod(ctx, cs, victim); err != nil {
-				logger.Error(err, "Preempted pod", "pod", klog.KObj(victim), "preemptor", klog.KObj(pod))
-				errCh.SendErrorWithCancel(err, cancel)
-				return
-			}
-			logger.V(2).Info("Preemptor Pod preempted victim Pod", "preemptor", klog.KObj(pod), "victim", klog.KObj(victim), "node", c.Name())
-		}
-
-		fh.EventRecorder().Eventf(victim, pod, v1.EventTypeNormal, "Preempted", "Preempting", "Preempted by pod %v on node %v", pod.UID, c.Name())
-	}
-
-	fh.Parallelizer().Until(ctx, len(c.Victims().Pods), preemptPod, ev.PluginName)
 	if err := errCh.ReceiveError(); err != nil {
 		return framework.AsStatus(err)
 	}
@@ -401,6 +460,91 @@ func (ev *Evaluator) prepareCandidate(ctx context.Context, c Candidate, pod *v1.
 	return nil
 }
 
+// prepareCandidateAsync triggers a goroutine for some preparation work:
+// - Evict the victim pods
+// - Reject the victim pods if they are in waitingPod map
+// - Clear the low-priority pods' nominatedNodeName status if needed
+// The Pod won't be retried until the goroutine triggered here completes.
+//
+// See http://kep.k8s.io/4832 for how the async preemption works.
+func (ev *Evaluator) prepareCandidateAsync(c Candidate, pod *v1.Pod, pluginName string) {
+	metrics.PreemptionVictims.Observe(float64(len(c.Victims().Pods)))
+
+	// Intentionally create a new context, not using a ctx from the scheduling cycle, to create ctx,
+	// because this process could continue even after this scheduling cycle finishes.
+	ctx, cancel := context.WithCancel(context.Background())
+	errCh := parallelize.NewErrorChannel()
+	preemptPod := func(index int) {
+		victim := c.Victims().Pods[index]
+		if err := ev.PreemptPod(ctx, c, pod, victim, pluginName); err != nil {
+			errCh.SendErrorWithCancel(err, cancel)
+		}
+	}
+
+	ev.mu.Lock()
+	ev.preempting.Insert(pod.UID)
+	ev.mu.Unlock()
+
+	logger := klog.FromContext(ctx)
+	go func() {
+		startTime := time.Now()
+		result := metrics.GoroutineResultSuccess
+		defer metrics.PreemptionGoroutinesDuration.WithLabelValues(result).Observe(metrics.SinceInSeconds(startTime))
+		defer metrics.PreemptionGoroutinesExecutionTotal.WithLabelValues(result).Inc()
+		defer func() {
+			if result == metrics.GoroutineResultError {
+				// When API call isn't successful, the Pod may get stuck in the unschedulable pod pool in the worst case.
+				// So, we should move the Pod to the activeQ.
+				ev.Handler.Activate(logger, map[string]*v1.Pod{pod.Name: pod})
+			}
+		}()
+		defer cancel()
+		logger.V(2).Info("Start the preemption asynchronously", "preemptor", klog.KObj(pod), "node", c.Name(), "numVictims", len(c.Victims().Pods))
+
+		// Lower priority pods nominated to run on this node, may no longer fit on
+		// this node. So, we should remove their nomination. Removing their
+		// nomination updates these pods and moves them to the active queue. It
+		// lets scheduler find another place for them.
+		nominatedPods := getLowerPriorityNominatedPods(logger, ev.Handler, pod, c.Name())
+		if err := util.ClearNominatedNodeName(ctx, ev.Handler.ClientSet(), nominatedPods...); err != nil {
+			logger.Error(err, "Cannot clear 'NominatedNodeName' field from lower priority pods on the same target node", "node", c.Name())
+			result = metrics.GoroutineResultError
+			// We do not return as this error is not critical.
+		}
+
+		if len(c.Victims().Pods) == 0 {
+			ev.mu.Lock()
+			delete(ev.preempting, pod.UID)
+			ev.mu.Unlock()
+
+			return
+		}
+
+		// We can evict all victims in parallel, but the last one.
+		// We have to remove the pod from the preempting map before the last one is evicted
+		// because, otherwise, the pod removal might be notified to the scheduling queue before
+		// we remove this pod from the preempting map,
+		// and the pod could end up stucking at the unschedulable pod pool
+		// by all the pod removal events being ignored.
+		ev.Handler.Parallelizer().Until(ctx, len(c.Victims().Pods)-1, preemptPod, ev.PluginName)
+		if err := errCh.ReceiveError(); err != nil {
+			logger.Error(err, "Error occurred during async preemption")
+			result = metrics.GoroutineResultError
+		}
+
+		ev.mu.Lock()
+		delete(ev.preempting, pod.UID)
+		ev.mu.Unlock()
+
+		if err := ev.PreemptPod(ctx, c, pod, c.Victims().Pods[len(c.Victims().Pods)-1], pluginName); err != nil {
+			logger.Error(err, "Error occurred during async preemption")
+			result = metrics.GoroutineResultError
+		}
+
+		logger.V(2).Info("Async Preemption finished completely", "preemptor", klog.KObj(pod), "node", c.Name(), "result", result)
+	}()
+}
+
 func getPodDisruptionBudgets(pdbLister policylisters.PodDisruptionBudgetLister) ([]*policy.PodDisruptionBudget, error) {
 	if pdbLister != nil {
 		return pdbLister.List(labels.Everything())
@@ -538,7 +682,7 @@ func getLowerPriorityNominatedPods(logger klog.Logger, pn framework.PodNominator
 // The number of candidates depends on the constraints defined in the plugin's args. In the returned list of
 // candidates, ones that do not violate PDB are preferred over ones that do.
 // NOTE: This method is exported for easier testing in default preemption.
-func (ev *Evaluator) DryRunPreemption(ctx context.Context, pod *v1.Pod, potentialNodes []*framework.NodeInfo,
+func (ev *Evaluator) DryRunPreemption(ctx context.Context, state *framework.CycleState, pod *v1.Pod, potentialNodes []*framework.NodeInfo,
 	pdbs []*policy.PodDisruptionBudget, offset int32, candidatesNum int32) ([]Candidate, *framework.NodeToStatus, error) {
 
 	fh := ev.Handler
@@ -557,7 +701,7 @@ func (ev *Evaluator) DryRunPreemption(ctx context.Context, pod *v1.Pod, potentia
 		nodeInfoCopy := potentialNodes[(int(offset)+i)%len(potentialNodes)].Snapshot()
 		logger.V(5).Info("Check the potential node for preemption", "node", nodeInfoCopy.Node().Name)
 
-		stateCopy := ev.State.Clone()
+		stateCopy := state.Clone()
 		pods, numPDBViolations, status := ev.SelectVictimsOnNode(ctx, stateCopy, pod, nodeInfoCopy, pdbs)
 		if status.IsSuccess() && len(pods) != 0 {
 			victims := extenderv1.Victims{

pkg/scheduler/framework/preemption/preemption_test.go

@@ -20,18 +20,26 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"reflect"
 	"sort"
+	"sync"
 	"testing"
+	"time"
 
 	"github.com/google/go-cmp/cmp"
 
 	v1 "k8s.io/api/core/v1"
 	policy "k8s.io/api/policy/v1"
 	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/apimachinery/pkg/types"
+	"k8s.io/apimachinery/pkg/util/sets"
+	"k8s.io/apimachinery/pkg/util/wait"
 	"k8s.io/client-go/informers"
 	clientsetfake "k8s.io/client-go/kubernetes/fake"
 	"k8s.io/client-go/kubernetes/scheme"
+	clienttesting "k8s.io/client-go/testing"
 	"k8s.io/client-go/tools/events"
+	"k8s.io/klog/v2"
 	"k8s.io/klog/v2/ktesting"
 	extenderv1 "k8s.io/kube-scheduler/extender/v1"
 	internalcache "k8s.io/kubernetes/pkg/scheduler/backend/cache"
@@ -81,6 +89,19 @@ func (pl *FakePostFilterPlugin) OrderedScoreFuncs(ctx context.Context, nodesToVi
 	return nil
 }
 
+type fakePodActivator struct {
+	activatedPods map[string]*v1.Pod
+	mu            *sync.RWMutex
+}
+
+func (f *fakePodActivator) Activate(logger klog.Logger, pods map[string]*v1.Pod) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	for name, pod := range pods {
+		f.activatedPods[name] = pod
+	}
+}
+
 type FakePreemptionScorePostFilterPlugin struct{}
 
 func (pl *FakePreemptionScorePostFilterPlugin) SelectVictimsOnNode(
@@ -243,9 +264,8 @@ func TestDryRunPreemption(t *testing.T) {
 					PluginName: "FakePostFilter",
 					Handler:    fwk,
 					Interface:  fakePostPlugin,
-					State:      state,
 				}
-				got, _, _ := pe.DryRunPreemption(ctx, pod, nodeInfos, nil, 0, int32(len(nodeInfos)))
+				got, _, _ := pe.DryRunPreemption(ctx, state, pod, nodeInfos, nil, 0, int32(len(nodeInfos)))
 				// Sort the values (inner victims) and the candidate itself (by its NominatedNodeName).
 				for i := range got {
 					victims := got[i].Victims().Pods
@@ -344,9 +364,8 @@ func TestSelectCandidate(t *testing.T) {
 					PluginName: "FakePreemptionScorePostFilter",
 					Handler:    fwk,
 					Interface:  fakePreemptionScorePostFilterPlugin,
-					State:      state,
 				}
-				candidates, _, _ := pe.DryRunPreemption(ctx, pod, nodeInfos, nil, 0, int32(len(nodeInfos)))
+				candidates, _, _ := pe.DryRunPreemption(ctx, state, pod, nodeInfos, nil, 0, int32(len(nodeInfos)))
 				s := pe.SelectCandidate(ctx, candidates)
 				if s == nil || len(s.Name()) == 0 {
 					t.Errorf("expect any node in %v, but no candidate selected", tt.expected)
@@ -393,6 +412,11 @@ func TestPrepareCandidate(t *testing.T) {
 			Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
 			Obj()
 
+		failVictim = st.MakePod().Name("fail-victim").UID("victim1").
+				Node(node1Name).SchedulerName(defaultSchedulerName).Priority(midPriority).
+				Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
+				Obj()
+
 		victim2 = st.MakePod().Name("victim2").UID("victim2").
 			Node(node1Name).SchedulerName(defaultSchedulerName).Priority(50000).
 			Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
@@ -404,6 +428,12 @@ func TestPrepareCandidate(t *testing.T) {
 						Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
 						Obj()
 
+		failVictim1WithMatchingCondition = st.MakePod().Name("fail-victim").UID("victim1").
+							Node(node1Name).SchedulerName(defaultSchedulerName).Priority(midPriority).
+							Conditions([]v1.PodCondition{condition}).
+							Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
+							Obj()
+
 		preemptor = st.MakePod().Name("preemptor").UID("preemptor").
 				SchedulerName(defaultSchedulerName).Priority(highPriority).
 				Containers([]v1.Container{st.MakeContainer().Name("container1").Obj()}).
@@ -416,10 +446,18 @@ func TestPrepareCandidate(t *testing.T) {
 		candidate             *fakeCandidate
 		preemptor             *v1.Pod
 		testPods              []*v1.Pod
+		expectedDeletedPods   []string
+		expectedDeletionError bool
+		expectedPatchError    bool
+		// Only compared when async preemption is disabled.
 		expectedStatus *framework.Status
+		// Only compared when async preemption is enabled.
+		expectedPreemptingMap sets.Set[types.UID]
+		expectedActivatedPods map[string]*v1.Pod
 	}{
 		{
 			name: "no victims",
+
 			candidate: &fakeCandidate{
 				victims: &extenderv1.Victims{},
 			},
@@ -429,9 +467,11 @@ func TestPrepareCandidate(t *testing.T) {
 			},
 			nodeNames:             []string{node1Name},
 			expectedStatus:        nil,
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
 		},
 		{
 			name: "one victim without condition",
+
 			candidate: &fakeCandidate{
 				name: node1Name,
 				victims: &extenderv1.Victims{
@@ -445,10 +485,13 @@ func TestPrepareCandidate(t *testing.T) {
 				victim1,
 			},
 			nodeNames:             []string{node1Name},
+			expectedDeletedPods:   []string{"victim1"},
 			expectedStatus:        nil,
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
 		},
 		{
 			name: "one victim with same condition",
+
 			candidate: &fakeCandidate{
 				name: node1Name,
 				victims: &extenderv1.Victims{
@@ -462,10 +505,13 @@ func TestPrepareCandidate(t *testing.T) {
 				victim1WithMatchingCondition,
 			},
 			nodeNames:             []string{node1Name},
+			expectedDeletedPods:   []string{"victim1"},
 			expectedStatus:        nil,
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
 		},
 		{
-			name: "one victim, but patch pod failed (not found victim1 pod)",
+			name: "one victim, not-found victim error is ignored when patching",
+
 			candidate: &fakeCandidate{
 				name: node1Name,
 				victims: &extenderv1.Victims{
@@ -477,10 +523,32 @@ func TestPrepareCandidate(t *testing.T) {
 			preemptor:             preemptor,
 			testPods:              []*v1.Pod{},
 			nodeNames:             []string{node1Name},
-			expectedStatus: framework.AsStatus(errors.New("patch pod status failed")),
+			expectedDeletedPods:   []string{"victim1"},
+			expectedStatus:        nil,
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
 		},
 		{
-			name: "one victim, but delete pod failed (not found victim1 pod)",
+			name: "one victim, but pod deletion failed",
+
+			candidate: &fakeCandidate{
+				name: node1Name,
+				victims: &extenderv1.Victims{
+					Pods: []*v1.Pod{
+						failVictim1WithMatchingCondition,
+					},
+				},
+			},
+			preemptor:             preemptor,
+			testPods:              []*v1.Pod{},
+			expectedDeletionError: true,
+			nodeNames:             []string{node1Name},
+			expectedStatus:        framework.AsStatus(errors.New("delete pod failed")),
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
+			expectedActivatedPods: map[string]*v1.Pod{preemptor.Name: preemptor},
+		},
+		{
+			name: "one victim, not-found victim error is ignored when deleting",
+
 			candidate: &fakeCandidate{
 				name: node1Name,
 				victims: &extenderv1.Victims{
@@ -492,15 +560,37 @@ func TestPrepareCandidate(t *testing.T) {
 			preemptor:             preemptor,
 			testPods:              []*v1.Pod{},
 			nodeNames:             []string{node1Name},
-			expectedStatus: framework.AsStatus(errors.New("delete pod failed")),
+			expectedDeletedPods:   []string{"victim1"},
+			expectedStatus:        nil,
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
 		},
 		{
-			name: "two victims without condition, one passes successfully and the second fails (not found victim2 pod)",
+			name: "one victim, but patch pod failed",
+
 			candidate: &fakeCandidate{
 				name: node1Name,
 				victims: &extenderv1.Victims{
 					Pods: []*v1.Pod{
-						victim1,
+						failVictim,
+					},
+				},
+			},
+			preemptor:             preemptor,
+			testPods:              []*v1.Pod{},
+			expectedPatchError:    true,
+			nodeNames:             []string{node1Name},
+			expectedStatus:        framework.AsStatus(errors.New("patch pod status failed")),
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
+			expectedActivatedPods: map[string]*v1.Pod{preemptor.Name: preemptor},
+		},
+		{
+			name: "two victims without condition, one passes successfully and the second fails",
+
+			candidate: &fakeCandidate{
+				name: node1Name,
+				victims: &extenderv1.Victims{
+					Pods: []*v1.Pod{
+						failVictim,
 						victim2,
 					},
 				},
@@ -510,12 +600,17 @@ func TestPrepareCandidate(t *testing.T) {
 				victim1,
 			},
 			nodeNames:             []string{node1Name},
+			expectedPatchError:    true,
+			expectedDeletedPods:   []string{"victim2"},
 			expectedStatus:        framework.AsStatus(errors.New("patch pod status failed")),
+			expectedPreemptingMap: sets.New(types.UID("preemptor")),
+			expectedActivatedPods: map[string]*v1.Pod{preemptor.Name: preemptor},
 		},
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%v (Async preemption enabled: %v)", tt.name, asyncPreemptionEnabled), func(t *testing.T) {
 				metrics.Register()
 				logger, ctx := ktesting.NewTestContext(t)
 				ctx, cancel := context.WithCancel(ctx)
@@ -533,9 +628,42 @@ func TestPrepareCandidate(t *testing.T) {
 				for _, pod := range tt.testPods {
 					objs = append(objs, pod)
 				}
+
+				requestStopper := make(chan struct{})
+				mu := &sync.RWMutex{}
+				deletedPods := sets.New[string]()
+				deletionFailure := false // whether any request to delete pod failed
+				patchFailure := false    // whether any request to patch pod status failed
+
 				cs := clientsetfake.NewClientset(objs...)
+				cs.PrependReactor("delete", "pods", func(action clienttesting.Action) (bool, runtime.Object, error) {
+					<-requestStopper
+					mu.Lock()
+					defer mu.Unlock()
+					name := action.(clienttesting.DeleteAction).GetName()
+					if name == "fail-victim" {
+						deletionFailure = true
+						return true, nil, fmt.Errorf("delete pod failed")
+					}
+
+					deletedPods.Insert(name)
+					return true, nil, nil
+				})
+
+				cs.PrependReactor("patch", "pods", func(action clienttesting.Action) (bool, runtime.Object, error) {
+					<-requestStopper
+					mu.Lock()
+					defer mu.Unlock()
+					if action.(clienttesting.PatchAction).GetName() == "fail-victim" {
+						patchFailure = true
+						return true, nil, fmt.Errorf("patch pod status failed")
+					}
+					return true, nil, nil
+				})
+
 				informerFactory := informers.NewSharedInformerFactory(cs, 0)
 				eventBroadcaster := events.NewBroadcaster(&events.EventSinkImpl{Interface: cs.EventsV1()})
+				fakeActivator := &fakePodActivator{activatedPods: make(map[string]*v1.Pod), mu: mu}
 				fwk, err := tf.NewFramework(
 					ctx,
 					registeredPlugins, "",
@@ -546,6 +674,7 @@ func TestPrepareCandidate(t *testing.T) {
 					frameworkruntime.WithSnapshotSharedLister(internalcache.NewSnapshot(tt.testPods, nodes)),
 					frameworkruntime.WithPodNominator(internalqueue.NewSchedulingQueue(nil, informerFactory)),
 					frameworkruntime.WithEventRecorder(eventBroadcaster.NewRecorder(scheme.Scheme, "test-scheduler")),
+					frameworkruntime.WithPodActivator(fakeActivator),
 				)
 				if err != nil {
 					t.Fatal(err)
@@ -553,13 +682,24 @@ func TestPrepareCandidate(t *testing.T) {
 				informerFactory.Start(ctx.Done())
 				informerFactory.WaitForCacheSync(ctx.Done())
 				fakePreemptionScorePostFilterPlugin := &FakePreemptionScorePostFilterPlugin{}
-			pe := Evaluator{
+				pe := NewEvaluator("FakePreemptionScorePostFilter", fwk, fakePreemptionScorePostFilterPlugin, asyncPreemptionEnabled)
-				PluginName: "FakePreemptionScorePostFilter",
-				Handler:    fwk,
-				Interface:  fakePreemptionScorePostFilterPlugin,
-				State:      framework.NewCycleState(),
-			}
 
+				if asyncPreemptionEnabled {
+					pe.prepareCandidateAsync(tt.candidate, tt.preemptor, "test-plugin")
+					pe.mu.Lock()
+					// The preempting map should be registered synchronously
+					// so we don't need wait.Poll.
+					if !tt.expectedPreemptingMap.Equal(pe.preempting) {
+						t.Errorf("expected preempting map %v, got %v", tt.expectedPreemptingMap, pe.preempting)
+						close(requestStopper)
+						pe.mu.Unlock()
+						return
+					}
+					pe.mu.Unlock()
+					// make the requests complete
+					close(requestStopper)
+				} else {
+					close(requestStopper) // no need to stop requests
 					status := pe.prepareCandidate(ctx, tt.candidate, tt.preemptor, "test-plugin")
 					if tt.expectedStatus == nil {
 						if status != nil {
@@ -570,11 +710,48 @@ func TestPrepareCandidate(t *testing.T) {
 							t.Errorf("expect status %v, but got nil", tt.expectedStatus)
 						} else if status.Code() != tt.expectedStatus.Code() {
 							t.Errorf("expect status code %v, but got %v", tt.expectedStatus.Code(), status.Code())
+						} else if status.Message() != tt.expectedStatus.Message() {
+							t.Errorf("expect status message %v, but got %v", tt.expectedStatus.Message(), status.Message())
 						}
 					}
+				}
+
+				var lastErrMsg string
+				if err := wait.PollUntilContextTimeout(ctx, time.Millisecond*200, wait.ForeverTestTimeout, false, func(ctx context.Context) (bool, error) {
+					mu.RLock()
+					defer mu.RUnlock()
+					if !deletedPods.Equal(sets.New(tt.expectedDeletedPods...)) {
+						lastErrMsg = fmt.Sprintf("expected deleted pods %v, got %v", tt.expectedDeletedPods, deletedPods.UnsortedList())
+						return false, nil
+					}
+					if tt.expectedDeletionError != deletionFailure {
+						lastErrMsg = fmt.Sprintf("expected deletion error %v, got %v", tt.expectedDeletionError, deletionFailure)
+						return false, nil
+					}
+					if tt.expectedPatchError != patchFailure {
+						lastErrMsg = fmt.Sprintf("expected patch error %v, got %v", tt.expectedPatchError, patchFailure)
+						return false, nil
+					}
+
+					if asyncPreemptionEnabled {
+						if tt.expectedActivatedPods != nil && !reflect.DeepEqual(tt.expectedActivatedPods, fakeActivator.activatedPods) {
+							lastErrMsg = fmt.Sprintf("expected activated pods %v, got %v", tt.expectedActivatedPods, fakeActivator.activatedPods)
+							return false, nil
+						}
+						if tt.expectedActivatedPods == nil && len(fakeActivator.activatedPods) != 0 {
+							lastErrMsg = fmt.Sprintf("expected no activated pods, got %v", fakeActivator.activatedPods)
+							return false, nil
+						}
+					}
+
+					return true, nil
+				}); err != nil {
+					t.Fatal(lastErrMsg)
+				}
 			})
 		}
 	}
+}
 
 type fakeExtender struct {
 	ignorable            bool
@@ -812,7 +989,6 @@ func TestCallExtenders(t *testing.T) {
 				PluginName: "FakePreemptionScorePostFilter",
 				Handler:    fwk,
 				Interface:  fakePreemptionScorePostFilterPlugin,
-				State:      framework.NewCycleState(),
 			}
 			gotCandidates, status := pe.callExtenders(logger, preemptor, tt.candidates)
 			if (tt.wantStatus == nil) != (status == nil) || status.Code() != tt.wantStatus.Code() {

pkg/scheduler/framework/runtime/framework.go

@@ -84,6 +84,7 @@ type frameworkImpl struct {
 
 	extenders []framework.Extender
 	framework.PodNominator
+	framework.PodActivator
 
 	parallelizer parallelize.Parallelizer
 }
@@ -131,6 +132,7 @@ type frameworkOptions struct {
 	snapshotSharedLister   framework.SharedLister
 	metricsRecorder        *metrics.MetricAsyncRecorder
 	podNominator           framework.PodNominator
+	podActivator           framework.PodActivator
 	extenders              []framework.Extender
 	captureProfile         CaptureProfile
 	parallelizer           parallelize.Parallelizer
@@ -200,6 +202,12 @@ func WithPodNominator(nominator framework.PodNominator) Option {
 	}
 }
 
+func WithPodActivator(activator framework.PodActivator) Option {
+	return func(o *frameworkOptions) {
+		o.podActivator = activator
+	}
+}
+
 // WithExtenders sets extenders for the scheduling frameworkImpl.
 func WithExtenders(extenders []framework.Extender) Option {
 	return func(o *frameworkOptions) {
@@ -279,6 +287,7 @@ func NewFramework(ctx context.Context, r Registry, profile *config.KubeScheduler
 		metricsRecorder:      options.metricsRecorder,
 		extenders:            options.extenders,
 		PodNominator:         options.podNominator,
+		PodActivator:         options.podActivator,
 		parallelizer:         options.parallelizer,
 		logger:               logger,
 	}
@@ -427,6 +436,10 @@ func (f *frameworkImpl) SetPodNominator(n framework.PodNominator) {
 	f.PodNominator = n
 }
 
+func (f *frameworkImpl) SetPodActivator(a framework.PodActivator) {
+	f.PodActivator = a
+}
+
 // Close closes each plugin, when they implement io.Closer interface.
 func (f *frameworkImpl) Close() error {
 	var errs []error

pkg/scheduler/framework/types.go

@@ -296,7 +296,7 @@ func (ce ClusterEvent) Label() string {
 
 // AllClusterEventLabels returns all possible cluster event labels given to the metrics.
 func AllClusterEventLabels() []string {
-	labels := []string{EventUnschedulableTimeout.Label()}
+	labels := []string{UnschedulableTimeout, ForceActivate}
 	for _, r := range allResources {
 		for _, a := range basicActionTypes {
 			labels = append(labels, ClusterEvent{Resource: r, ActionType: a}.Label())

pkg/scheduler/metrics/metrics.go

@@ -40,6 +40,11 @@ const (
 	Binding = "binding"
 )
 
+const (
+	GoroutineResultSuccess = "success"
+	GoroutineResultError   = "error"
+)
+
 // ExtentionPoints is a list of possible values for the extension_point label.
 var ExtentionPoints = []string{
 	PreFilter,
@@ -105,13 +110,20 @@ var (
 	FrameworkExtensionPointDuration *metrics.HistogramVec
 	PluginExecutionDuration         *metrics.HistogramVec
 
-	// This is only available when the QHint feature gate is enabled.
-	queueingHintExecutionDuration *metrics.HistogramVec
-	SchedulerQueueIncomingPods    *metrics.CounterVec
 	PermitWaitDuration    *metrics.HistogramVec
 	CacheSize             *metrics.GaugeVec
 	unschedulableReasons  *metrics.GaugeVec
 	PluginEvaluationTotal *metrics.CounterVec
+
+	// The below two are only available when the QHint feature gate is enabled.
+	queueingHintExecutionDuration *metrics.HistogramVec
+	SchedulerQueueIncomingPods    *metrics.CounterVec
+
+	// The below two are only available when the async-preemption feature gate is enabled.
+	PreemptionGoroutinesDuration       *metrics.HistogramVec
+	PreemptionGoroutinesExecutionTotal *metrics.CounterVec
+
+	// metricsList is a list of all metrics that should be registered always, regardless of any feature gate's value.
 	metricsList []metrics.Registerable
 )
 
@@ -123,11 +135,14 @@ func Register() {
 	registerMetrics.Do(func() {
 		InitMetrics()
 		RegisterMetrics(metricsList...)
-		if utilfeature.DefaultFeatureGate.Enabled(features.SchedulerQueueingHints) {
-			RegisterMetrics(queueingHintExecutionDuration)
-			RegisterMetrics(InFlightEvents)
-		}
 		volumebindingmetrics.RegisterVolumeSchedulingMetrics()
+
+		if utilfeature.DefaultFeatureGate.Enabled(features.SchedulerQueueingHints) {
+			RegisterMetrics(queueingHintExecutionDuration, InFlightEvents)
+		}
+		if utilfeature.DefaultFeatureGate.Enabled(features.SchedulerAsyncPreemption) {
+			RegisterMetrics(PreemptionGoroutinesDuration, PreemptionGoroutinesExecutionTotal)
+		}
 	})
 }
 
@@ -317,6 +332,25 @@ func InitMetrics() {
 			StabilityLevel: metrics.ALPHA,
 		}, []string{"plugin", "extension_point", "profile"})
 
+	PreemptionGoroutinesDuration = metrics.NewHistogramVec(
+		&metrics.HistogramOpts{
+			Subsystem:      SchedulerSubsystem,
+			Name:           "preemption_goroutines_duration_seconds",
+			Help:           "Duration in seconds for running goroutines for the preemption.",
+			Buckets:        metrics.ExponentialBuckets(0.01, 2, 20),
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{"result"})
+
+	PreemptionGoroutinesExecutionTotal = metrics.NewCounterVec(
+		&metrics.CounterOpts{
+			Subsystem:      SchedulerSubsystem,
+			Name:           "preemption_goroutines_execution_total",
+			Help:           "Number of preemption goroutines executed.",
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{"result"})
+
 	metricsList = []metrics.Registerable{
 		scheduleAttempts,
 		schedulingLatency,

pkg/scheduler/scheduler.go

@@ -355,6 +355,7 @@ func New(ctx context.Context,
 
 	for _, fwk := range profiles {
 		fwk.SetPodNominator(podQueue)
+		fwk.SetPodActivator(podQueue)
 	}
 
 	schedulerCache := internalcache.New(ctx, durationToExpireAssumedPod)

pkg/scheduler/testing/wrappers.go

@@ -310,6 +310,12 @@ func (p *PodWrapper) Name(s string) *PodWrapper {
 	return p
 }
 
+// Name sets `s` as the name of the inner pod.
+func (p *PodWrapper) GenerateName(s string) *PodWrapper {
+	p.SetGenerateName(s)
+	return p
+}
+
 // UID sets `s` as the UID of the inner pod.
 func (p *PodWrapper) UID(s string) *PodWrapper {
 	p.SetUID(types.UID(s))

test/e2e/feature/feature.go

@@ -312,6 +312,10 @@ var (
 	// TODO: document the feature (owning SIG, when to use this feature for a test)
 	RegularResourceUsageTracking = framework.WithFeature(framework.ValidFeatures.Add("RegularResourceUsageTracking"))
 
+	// Owner: sig-scheduling
+	// Marks tests of the asynchronous preemption (KEP-4832) that require the `SchedulerAsyncPreemption` feature gate.
+	SchedulerAsyncPreemption = framework.WithFeature(framework.ValidFeatures.Add("SchedulerAsyncPreemption"))
+
 	// Owner: sig-network
 	// Marks tests that require a pod networking implementation that supports SCTP
 	// traffic between pods.

test/e2e/scheduling/preemption.go

@@ -43,6 +43,7 @@ import (
 	clientset "k8s.io/client-go/kubernetes"
 	"k8s.io/client-go/tools/cache"
 	"k8s.io/kubernetes/pkg/apis/scheduling"
+	"k8s.io/kubernetes/test/e2e/feature"
 	"k8s.io/kubernetes/test/e2e/framework"
 	e2enode "k8s.io/kubernetes/test/e2e/framework/node"
 	e2epod "k8s.io/kubernetes/test/e2e/framework/pod"
@@ -305,6 +306,157 @@ var _ = SIGDescribe("SchedulerPreemption", framework.WithSerial(), func() {
 		}
 	})
 
+	/*
+		Release: v1.32
+		Testname: Scheduler runs the preemption with various priority classes expectedly
+		Description: When there are Pods with various priority classes running the preemption,
+		the scheduler must prioritize the Pods with the higher priority class.
+	*/
+	framework.It("validates various priority Pods preempt expectedly with the async preemption", feature.SchedulerAsyncPreemption, func(ctx context.Context) {
+		var podRes v1.ResourceList
+		// Create 10 pods per node that will eat up all the node's resources.
+		ginkgo.By("Create 10 low-priority pods on each node.")
+		lowPriorityPods := make([]*v1.Pod, 0, 10*len(nodeList.Items))
+		// Create pods in the cluster.
+		for i, node := range nodeList.Items {
+			// Update each node to advertise 3 available extended resources
+			e2enode.AddExtendedResource(ctx, cs, node.Name, testExtendedResource, resource.MustParse("10"))
+
+			// Create 10 low priority pods on each node, which will use up 10/10 of the node's resources.
+			for j := 0; j < 10; j++ {
+				// Request 1 of the available resources for the victim pods
+				podRes = v1.ResourceList{}
+				podRes[testExtendedResource] = resource.MustParse("1")
+				pausePod := createPausePod(ctx, f, pausePodConfig{
+					Name:              fmt.Sprintf("pod%d-%d-%v", i, j, lowPriorityClassName),
+					PriorityClassName: lowPriorityClassName,
+					// This victim pod will be preempted by the high priority pod.
+					// But, the deletion will be blocked by the finalizer.
+					//
+					// The finalizer is needed to prevent the medium Pods from being scheduled instead of the high Pods,
+					// depending on when the scheduler notices the existence of all the high Pods we create.
+					Finalizers: []string{testFinalizer},
+					Resources: &v1.ResourceRequirements{
+						Requests: podRes,
+						Limits:   podRes,
+					},
+					Affinity: &v1.Affinity{
+						NodeAffinity: &v1.NodeAffinity{
+							RequiredDuringSchedulingIgnoredDuringExecution: &v1.NodeSelector{
+								NodeSelectorTerms: []v1.NodeSelectorTerm{
+									{
+										MatchFields: []v1.NodeSelectorRequirement{
+											{Key: "metadata.name", Operator: v1.NodeSelectorOpIn, Values: []string{node.Name}},
+										},
+									},
+								},
+							},
+						},
+					},
+				})
+				lowPriorityPods = append(lowPriorityPods, pausePod)
+				framework.Logf("Created pod: %v", pausePod.Name)
+			}
+		}
+
+		ginkgo.By("Wait for lower priority pods to be scheduled.")
+		for _, pod := range lowPriorityPods {
+			framework.ExpectNoError(e2epod.WaitForPodRunningInNamespace(ctx, cs, pod))
+		}
+
+		highPriorityPods := make([]*v1.Pod, 0, 5*len(nodeList.Items))
+		mediumPriorityPods := make([]*v1.Pod, 0, 10*len(nodeList.Items))
+
+		ginkgo.By("Run high/medium priority pods that have same requirements as that of lower priority pod")
+		for i := range nodeList.Items {
+			// Create medium priority pods first
+			// to confirm the scheduler finally prioritize the high priority pods, ignoring the medium priority pods.
+			for j := 0; j < 10; j++ {
+				// 5 pods per node will be unschedulable
+				// because the node only has 10 resource, and high priority pods will use 5 resource.
+				p := createPausePod(ctx, f, pausePodConfig{
+					Name:              fmt.Sprintf("pod%d-%d-%v", i, j, mediumPriorityClassName),
+					PriorityClassName: mediumPriorityClassName,
+					Resources: &v1.ResourceRequirements{
+						// Set the pod request to the low priority pod's resources
+						Requests: lowPriorityPods[0].Spec.Containers[0].Resources.Requests,
+						Limits:   lowPriorityPods[0].Spec.Containers[0].Resources.Requests,
+					},
+				})
+				mediumPriorityPods = append(mediumPriorityPods, p)
+			}
+
+			for j := 0; j < 5; j++ {
+				p := createPausePod(ctx, f, pausePodConfig{
+					Name:              fmt.Sprintf("pod%d-%d-%v", i, j, highPriorityClassName),
+					PriorityClassName: highPriorityClassName,
+					Resources: &v1.ResourceRequirements{
+						// Set the pod request to the low priority pod's resources
+						Requests: lowPriorityPods[0].Spec.Containers[0].Resources.Requests,
+						Limits:   lowPriorityPods[0].Spec.Containers[0].Resources.Requests,
+					},
+				})
+				highPriorityPods = append(highPriorityPods, p)
+			}
+		}
+
+		// All low priority Pods should be the target of preemption.
+		// Those Pods have a finalizer and hence should not be deleted yet at this point.
+		ginkgo.By("Check all low priority pods to be about to preempted.")
+		for _, pod := range lowPriorityPods {
+			framework.ExpectNoError(wait.PollUntilContextTimeout(ctx, time.Second, framework.PodStartTimeout, false, func(ctx context.Context) (bool, error) {
+				preemptedPod, err := cs.CoreV1().Pods(pod.Namespace).Get(ctx, pod.Name, metav1.GetOptions{})
+				if err != nil {
+					return false, err
+				}
+				return preemptedPod.DeletionTimestamp != nil, nil
+			}))
+		}
+
+		// All high priority Pods should be schedulable by removing the low priority Pods.
+		ginkgo.By("Wait for high priority pods to be ready for the preemption.")
+		for _, pod := range highPriorityPods {
+			framework.ExpectNoError(wait.PollUntilContextTimeout(ctx, time.Second, framework.PodStartTimeout, false, func(ctx context.Context) (bool, error) {
+				highPod, err := cs.CoreV1().Pods(pod.Namespace).Get(ctx, pod.Name, metav1.GetOptions{})
+				if err != nil {
+					return false, err
+				}
+				return highPod.Status.NominatedNodeName != "", nil
+			}))
+		}
+
+		ginkgo.By("Remove the finalizer from all low priority pods to proceed the preemption.")
+		for _, pod := range lowPriorityPods {
+			// Remove the finalizer so that the pod can be deleted by GC
+			e2epod.NewPodClient(f).RemoveFinalizer(ctx, pod.Name, testFinalizer)
+		}
+
+		ginkgo.By("Wait for high priority pods to be scheduled.")
+		for _, pod := range highPriorityPods {
+			framework.ExpectNoError(e2epod.WaitForPodRunningInNamespace(ctx, cs, pod))
+		}
+
+		ginkgo.By("Wait for 5 medium priority pods to be scheduled.")
+		framework.ExpectNoError(wait.PollUntilContextTimeout(ctx, time.Second, framework.PodStartTimeout, false, func(ctx context.Context) (bool, error) {
+			scheduled := 0
+			for _, pod := range mediumPriorityPods {
+				medPod, err := cs.CoreV1().Pods(pod.Namespace).Get(ctx, pod.Name, metav1.GetOptions{})
+				if err != nil {
+					return false, err
+				}
+
+				if medPod.Spec.NodeName != "" {
+					scheduled++
+				}
+			}
+			if scheduled > 5 {
+				return false, fmt.Errorf("expected 5 medium priority pods to be scheduled, but got %d", scheduled)
+			}
+
+			return scheduled == 5, nil
+		}))
+	})
+
 	/*
 		Release: v1.31
 		Testname: Verify the DisruptionTarget condition is added to the preempted pod

test/featuregates_linter/test_data/versioned_feature_list.yaml

@@ -1058,6 +1058,12 @@
     lockToDefault: false
     preRelease: Alpha
     version: "1.29"
+- name: SchedulerAsyncPreemption
+  versionedSpecs:
+  - default: false
+    lockToDefault: false
+    preRelease: Alpha
+    version: "1.32"
 - name: SchedulerQueueingHints
   versionedSpecs:
   - default: false

test/integration/scheduler/preemption/preemption_test.go

@@ -33,23 +33,34 @@ import (
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/apimachinery/pkg/util/intstr"
 	"k8s.io/apimachinery/pkg/util/wait"
+	utilfeature "k8s.io/apiserver/pkg/util/feature"
 	"k8s.io/client-go/informers"
 	clientset "k8s.io/client-go/kubernetes"
 	restclient "k8s.io/client-go/rest"
+	featuregatetesting "k8s.io/component-base/featuregate/testing"
 	"k8s.io/component-helpers/storage/volume"
 	"k8s.io/klog/v2"
 	configv1 "k8s.io/kube-scheduler/config/v1"
 	podutil "k8s.io/kubernetes/pkg/api/v1/pod"
 	"k8s.io/kubernetes/pkg/apis/scheduling"
+	"k8s.io/kubernetes/pkg/features"
 	"k8s.io/kubernetes/pkg/scheduler"
+	"k8s.io/kubernetes/pkg/scheduler/apis/config"
 	configtesting "k8s.io/kubernetes/pkg/scheduler/apis/config/testing"
+	"k8s.io/kubernetes/pkg/scheduler/backend/queue"
 	"k8s.io/kubernetes/pkg/scheduler/framework"
+	"k8s.io/kubernetes/pkg/scheduler/framework/plugins/defaultpreemption"
+	plfeature "k8s.io/kubernetes/pkg/scheduler/framework/plugins/feature"
+	"k8s.io/kubernetes/pkg/scheduler/framework/plugins/names"
 	"k8s.io/kubernetes/pkg/scheduler/framework/plugins/volumerestrictions"
+	"k8s.io/kubernetes/pkg/scheduler/framework/preemption"
 	frameworkruntime "k8s.io/kubernetes/pkg/scheduler/framework/runtime"
 	st "k8s.io/kubernetes/pkg/scheduler/testing"
 	"k8s.io/kubernetes/plugin/pkg/admission/priority"
 	testutils "k8s.io/kubernetes/test/integration/util"
+	"k8s.io/kubernetes/test/utils/ktesting"
 	"k8s.io/utils/pointer"
+	"k8s.io/utils/ptr"
 )
 
 // imported from testutils
@@ -452,8 +463,11 @@ func TestPreemption(t *testing.T) {
 		t.Fatalf("Error creating node: %v", err)
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
+				featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.SchedulerAsyncPreemption, asyncPreemptionEnabled)
+
 				filter.Tokens = test.initTokens
 				filter.EnablePreFilter = test.enablePreFilter
 				filter.Unresolvable = test.unresolvable
@@ -485,12 +499,10 @@ func TestPreemption(t *testing.T) {
 						if cond == nil {
 							t.Errorf("Pod %q does not have the expected condition: %q", klog.KObj(pod), v1.DisruptionTarget)
 						}
-				} else {
+					} else if p.DeletionTimestamp != nil {
-					if p.DeletionTimestamp != nil {
 						t.Errorf("Didn't expect pod %v to get preempted.", p.Name)
 					}
 				}
-			}
 				// Also check that the preemptor pod gets the NominatedNodeName field set.
 				if len(test.preemptedPodIndexes) > 0 {
 					if err := waitForNominatedNodeName(cs, preemptor); err != nil {
@@ -504,6 +516,508 @@ func TestPreemption(t *testing.T) {
 			})
 		}
 	}
+}
+
+func TestAsyncPreemption(t *testing.T) {
+	type createPod struct {
+		pod *v1.Pod
+		// count is the number of times the pod should be created by this action.
+		// i.e., if you use it, you have to use GenerateName.
+		// By default, it's 1.
+		count *int
+	}
+
+	type schedulePod struct {
+		podName       string
+		expectSuccess bool
+	}
+
+	type scenario struct {
+		// name is this step's name, just for the debugging purpose.
+		name string
+
+		// Only one of the following actions should be set.
+
+		// createPod creates a Pod.
+		createPod *createPod
+		// schedulePod schedules one Pod that is at the top of the activeQ.
+		// You should give a Pod name that is supposed to be scheduled.
+		schedulePod *schedulePod
+		// completePreemption completes the preemption that is currently on-going.
+		// You should give a Pod name.
+		completePreemption string
+		// podGatedInQueue checks if the given Pod is in the scheduling queue and gated by the preemption.
+		// You should give a Pod name.
+		podGatedInQueue string
+		// podRunningPreemption checks if the given Pod is running preemption.
+		// You should give a Pod index representing the order of Pod creation.
+		// e.g., if you want to check the Pod created first in the test case, you should give 0.
+		podRunningPreemption *int
+	}
+
+	tests := []struct {
+		name string
+		// scenarios after the first attempt of scheduling the pod.
+		scenarios []scenario
+	}{
+		{
+			// Very basic test case: if it fails, the basic scenario is broken somewhere.
+			name: "basic: async preemption happens expectedly",
+			scenarios: []scenario{
+				{
+					name: "create scheduled Pod",
+					createPod: &createPod{
+						pod:   st.MakePod().GenerateName("victim-").Req(map[v1.ResourceName]string{v1.ResourceCPU: "2"}).Node("node").Container("image").ZeroTerminationGracePeriod().Priority(1).Obj(),
+						count: ptr.To(2),
+					},
+				},
+				{
+					name: "create a preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor").Req(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Container("image").Priority(100).Obj(),
+					},
+				},
+				{
+					name: "schedule the preemptor Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor",
+					},
+				},
+				{
+					name:            "check the pod is in the queue and gated",
+					podGatedInQueue: "preemptor",
+				},
+				{
+					name:                 "check the preemptor Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(2),
+				},
+				{
+					name:               "complete the preemption API calls",
+					completePreemption: "preemptor",
+				},
+				{
+					name: "schedule the preemptor Pod after the preemption",
+					schedulePod: &schedulePod{
+						podName:       "preemptor",
+						expectSuccess: true,
+					},
+				},
+			},
+		},
+		{
+			name: "Lower priority Pod doesn't take over the place for higher priority Pod that is running the preemption",
+			scenarios: []scenario{
+				{
+					name: "create scheduled Pod",
+					createPod: &createPod{
+						pod:   st.MakePod().GenerateName("victim-").Req(map[v1.ResourceName]string{v1.ResourceCPU: "2"}).Node("node").Container("image").ZeroTerminationGracePeriod().Priority(1).Obj(),
+						count: ptr.To(2),
+					},
+				},
+				{
+					name: "create a preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor-high-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Container("image").Priority(100).Obj(),
+					},
+				},
+				{
+					name: "schedule the preemptor Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-high-priority",
+					},
+				},
+				{
+					name:            "check the pod is in the queue and gated",
+					podGatedInQueue: "preemptor-high-priority",
+				},
+				{
+					name:                 "check the preemptor Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(2),
+				},
+				{
+					// This Pod is lower priority than the preemptor Pod.
+					// Given the preemptor Pod is nominated to the node, this Pod should be unschedulable.
+					name: "create a second Pod that is lower priority than the first preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("pod-mid-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Container("image").Priority(50).Obj(),
+					},
+				},
+				{
+					name: "schedule the mid-priority Pod",
+					schedulePod: &schedulePod{
+						podName: "pod-mid-priority",
+					},
+				},
+				{
+					name:               "complete the preemption API calls",
+					completePreemption: "preemptor-high-priority",
+				},
+				{
+					// the preemptor pod should be popped from the queue before the mid-priority pod.
+					name: "schedule the preemptor Pod again",
+					schedulePod: &schedulePod{
+						podName:       "preemptor-high-priority",
+						expectSuccess: true,
+					},
+				},
+				{
+					name: "schedule the mid-priority Pod again",
+					schedulePod: &schedulePod{
+						podName: "pod-mid-priority",
+					},
+				},
+			},
+		},
+		{
+			name: "Higher priority Pod takes over the place for lower priority Pod that is running the preemption",
+			scenarios: []scenario{
+				{
+					name: "create scheduled Pod",
+					createPod: &createPod{
+						pod:   st.MakePod().GenerateName("victim-").Req(map[v1.ResourceName]string{v1.ResourceCPU: "1"}).Node("node").Container("image").ZeroTerminationGracePeriod().Priority(1).Obj(),
+						count: ptr.To(4),
+					},
+				},
+				{
+					name: "create a preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor-high-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Container("image").Priority(100).Obj(),
+					},
+				},
+				{
+					name: "schedule the preemptor Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-high-priority",
+					},
+				},
+				{
+					name:            "check the pod is in the queue and gated",
+					podGatedInQueue: "preemptor-high-priority",
+				},
+				{
+					name:                 "check the preemptor Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(4),
+				},
+				{
+					// This Pod is higher priority than the preemptor Pod.
+					// Even though the preemptor Pod is nominated to the node, this Pod can take over the place.
+					name: "create a second Pod that is higher priority than the first preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor-super-high-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Container("image").Priority(200).Obj(),
+					},
+				},
+				{
+					name: "schedule the super-high-priority Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-super-high-priority",
+					},
+				},
+				{
+					name:                 "check the super-high-priority Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(5),
+				},
+				{
+					// the super-high-priority preemptor should enter the preemption
+					// and select the place where the preemptor-high-priority selected.
+					// So, basically both goroutines are preempting the same Pods.
+					name:            "check the super-high-priority pod is in the queue and gated",
+					podGatedInQueue: "preemptor-super-high-priority",
+				},
+				{
+					name:               "complete the preemption API calls of super-high-priority",
+					completePreemption: "preemptor-super-high-priority",
+				},
+				{
+					name:               "complete the preemption API calls of high-priority",
+					completePreemption: "preemptor-high-priority",
+				},
+				{
+					name: "schedule the super-high-priority Pod",
+					schedulePod: &schedulePod{
+						podName:       "preemptor-super-high-priority",
+						expectSuccess: true,
+					},
+				},
+				{
+					name: "schedule the high-priority Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-high-priority",
+					},
+				},
+			},
+		},
+		{
+			name: "Lower priority Pod can select the same place where the higher priority Pod is preempting if the node is big enough",
+			scenarios: []scenario{
+				{
+					name: "create scheduled Pod",
+					createPod: &createPod{
+						pod:   st.MakePod().GenerateName("victim-").Req(map[v1.ResourceName]string{v1.ResourceCPU: "1"}).Node("node").Container("image").ZeroTerminationGracePeriod().Priority(1).Obj(),
+						count: ptr.To(4),
+					},
+				},
+				{
+					// It will preempt two victims.
+					name: "create a preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor-high-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "2"}).Container("image").Priority(100).Obj(),
+					},
+				},
+				{
+					name: "schedule the preemptor Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-high-priority",
+					},
+				},
+				{
+					name:            "check the pod is in the queue and gated",
+					podGatedInQueue: "preemptor-high-priority",
+				},
+				{
+					name:                 "check the preemptor Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(4),
+				},
+				{
+					// This Pod is lower priority than the preemptor Pod.
+					// Given the preemptor Pod is nominated to the node, this Pod should be unschedulable.
+					// This Pod will trigger the preemption to target the two victims that the first Pod doesn't target.
+					name: "create a second Pod that is lower priority than the first preemptor Pod",
+					createPod: &createPod{
+						pod: st.MakePod().Name("preemptor-mid-priority").Req(map[v1.ResourceName]string{v1.ResourceCPU: "2"}).Container("image").Priority(50).Obj(),
+					},
+				},
+				{
+					name: "schedule the mid-priority Pod",
+					schedulePod: &schedulePod{
+						podName: "preemptor-mid-priority",
+					},
+				},
+				{
+					name:            "check the mid-priority pod is in the queue and gated",
+					podGatedInQueue: "preemptor-mid-priority",
+				},
+				{
+					name:                 "check the mid-priority Pod making the preemption API calls",
+					podRunningPreemption: ptr.To(5),
+				},
+				{
+					name:               "complete the preemption API calls",
+					completePreemption: "preemptor-mid-priority",
+				},
+				{
+					name:               "complete the preemption API calls",
+					completePreemption: "preemptor-high-priority",
+				},
+				{
+					// the preemptor pod should be popped from the queue before the mid-priority pod.
+					name: "schedule the preemptor Pod again",
+					schedulePod: &schedulePod{
+						podName:       "preemptor-high-priority",
+						expectSuccess: true,
+					},
+				},
+				{
+					name: "schedule the mid-priority Pod again",
+					schedulePod: &schedulePod{
+						podName:       "preemptor-mid-priority",
+						expectSuccess: true,
+					},
+				},
+			},
+		},
+	}
+
+	// All test cases have the same node.
+	node := st.MakeNode().Name("node").Capacity(map[v1.ResourceName]string{v1.ResourceCPU: "4"}).Obj()
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			// We need to use a custom preemption plugin to test async preemption behavior
+			delayedPreemptionPluginName := "delay-preemption"
+			// keyed by the pod name
+			preemptionDoneChannels := make(map[string]chan struct{})
+			defer func() {
+				for _, ch := range preemptionDoneChannels {
+					close(ch)
+				}
+			}()
+			registry := make(frameworkruntime.Registry)
+			var preemptionPlugin *defaultpreemption.DefaultPreemption
+			err := registry.Register(delayedPreemptionPluginName, func(c context.Context, r runtime.Object, fh framework.Handle) (framework.Plugin, error) {
+				p, err := frameworkruntime.FactoryAdapter(plfeature.Features{EnableAsyncPreemption: true}, defaultpreemption.New)(c, &config.DefaultPreemptionArgs{
+					// Set default values to pass the validation at the initialization, not related to the test.
+					MinCandidateNodesPercentage: 10,
+					MinCandidateNodesAbsolute:   100,
+				}, fh)
+				if err != nil {
+					return nil, fmt.Errorf("error creating default preemption plugin: %w", err)
+				}
+
+				var ok bool
+				preemptionPlugin, ok = p.(*defaultpreemption.DefaultPreemption)
+				if !ok {
+					return nil, fmt.Errorf("unexpected plugin type %T", p)
+				}
+
+				preemptPodFn := preemptionPlugin.Evaluator.PreemptPod
+				preemptionPlugin.Evaluator.PreemptPod = func(ctx context.Context, c preemption.Candidate, preemptor, victim *v1.Pod, pluginName string) error {
+					// block the preemption goroutine to complete until the test case allows it to proceed.
+					if ch, ok := preemptionDoneChannels[preemptor.Name]; ok {
+						<-ch
+					}
+					return preemptPodFn(ctx, c, preemptor, victim, pluginName)
+				}
+
+				return preemptionPlugin, nil
+			})
+			if err != nil {
+				t.Fatalf("Error registering a filter: %v", err)
+			}
+			cfg := configtesting.V1ToInternalWithDefaults(t, configv1.KubeSchedulerConfiguration{
+				Profiles: []configv1.KubeSchedulerProfile{{
+					SchedulerName: pointer.String(v1.DefaultSchedulerName),
+					Plugins: &configv1.Plugins{
+						MultiPoint: configv1.PluginSet{
+							Enabled: []configv1.Plugin{
+								{Name: delayedPreemptionPluginName},
+							},
+							Disabled: []configv1.Plugin{
+								{Name: names.DefaultPreemption},
+							},
+						},
+					},
+				}},
+			})
+
+			// It initializes the scheduler, but doesn't start.
+			// We manually trigger the scheduling cycle.
+			testCtx := testutils.InitTestSchedulerWithOptions(t,
+				testutils.InitTestAPIServer(t, "preemption", nil),
+				0,
+				scheduler.WithProfiles(cfg.Profiles...),
+				scheduler.WithFrameworkOutOfTreeRegistry(registry),
+				// disable backoff
+				scheduler.WithPodMaxBackoffSeconds(0),
+				scheduler.WithPodInitialBackoffSeconds(0),
+			)
+			testutils.SyncSchedulerInformerFactory(testCtx)
+			cs := testCtx.ClientSet
+
+			if preemptionPlugin == nil {
+				t.Fatalf("the preemption plugin should be initialized")
+			}
+
+			logger, _ := ktesting.NewTestContext(t)
+			featuregatetesting.SetFeatureGateDuringTest(t, utilfeature.DefaultFeatureGate, features.SchedulerAsyncPreemption, true)
+
+			createdPods := []*v1.Pod{}
+			defer testutils.CleanupPods(testCtx.Ctx, cs, t, createdPods)
+
+			ctx, cancel := context.WithCancel(context.Background())
+			defer cancel()
+
+			if _, err := cs.CoreV1().Nodes().Create(ctx, node, metav1.CreateOptions{}); err != nil {
+				t.Fatalf("Failed to create an initial Node %q: %v", node.Name, err)
+			}
+			defer func() {
+				if err := cs.CoreV1().Nodes().Delete(ctx, node.Name, metav1.DeleteOptions{}); err != nil {
+					t.Fatalf("Failed to delete the Node %q: %v", node.Name, err)
+				}
+			}()
+
+			for _, scenario := range test.scenarios {
+				t.Logf("Running scenario: %s", scenario.name)
+				switch {
+				case scenario.createPod != nil:
+					if scenario.createPod.count == nil {
+						scenario.createPod.count = ptr.To(1)
+					}
+
+					for i := 0; i < *scenario.createPod.count; i++ {
+						pod, err := cs.CoreV1().Pods(testCtx.NS.Name).Create(ctx, scenario.createPod.pod, metav1.CreateOptions{})
+						if err != nil {
+							t.Fatalf("Failed to create a Pod %q: %v", pod.Name, err)
+						}
+						createdPods = append(createdPods, pod)
+					}
+				case scenario.schedulePod != nil:
+					lastFailure := ""
+					if err := wait.PollUntilContextTimeout(testCtx.Ctx, time.Millisecond*200, wait.ForeverTestTimeout, false, func(ctx context.Context) (bool, error) {
+						if len(testCtx.Scheduler.SchedulingQueue.PodsInActiveQ()) == 0 {
+							lastFailure = fmt.Sprintf("Expected the pod %s to be scheduled, but no pod arrives at the activeQ", scenario.schedulePod.podName)
+							return false, nil
+						}
+
+						if testCtx.Scheduler.SchedulingQueue.PodsInActiveQ()[0].Name != scenario.schedulePod.podName {
+							// need to wait more because maybe the queue will get another Pod that higher priority than the current top pod.
+							lastFailure = fmt.Sprintf("The pod %s is expected to be scheduled, but the top Pod is %s", scenario.schedulePod.podName, testCtx.Scheduler.SchedulingQueue.PodsInActiveQ()[0].Name)
+							return false, nil
+						}
+
+						return true, nil
+					}); err != nil {
+						t.Fatal(lastFailure)
+					}
+
+					preemptionDoneChannels[scenario.schedulePod.podName] = make(chan struct{})
+					testCtx.Scheduler.ScheduleOne(testCtx.Ctx)
+					if scenario.schedulePod.expectSuccess {
+						if err := wait.PollUntilContextTimeout(testCtx.Ctx, 200*time.Millisecond, wait.ForeverTestTimeout, false, testutils.PodScheduled(cs, testCtx.NS.Name, scenario.schedulePod.podName)); err != nil {
+							t.Fatalf("Expected the pod %s to be scheduled", scenario.schedulePod.podName)
+						}
+					} else {
+						if !podInUnschedulablePodPool(t, testCtx.Scheduler.SchedulingQueue, scenario.schedulePod.podName) {
+							t.Fatalf("Expected the pod %s to be in the queue after the scheduling attempt", scenario.schedulePod.podName)
+						}
+					}
+				case scenario.completePreemption != "":
+					if _, ok := preemptionDoneChannels[scenario.completePreemption]; !ok {
+						t.Fatalf("The preemptor Pod %q is not running preemption", scenario.completePreemption)
+					}
+
+					close(preemptionDoneChannels[scenario.completePreemption])
+					delete(preemptionDoneChannels, scenario.completePreemption)
+				case scenario.podGatedInQueue != "":
+					// make sure the Pod is in the queue in the first place.
+					if !podInUnschedulablePodPool(t, testCtx.Scheduler.SchedulingQueue, scenario.podGatedInQueue) {
+						t.Fatalf("Expected the pod %s to be in the queue", scenario.podGatedInQueue)
+					}
+
+					// Make sure this Pod is gated by the preemption at PreEnqueue extension point
+					// by activating the Pod and see if it's still in the unsched pod pool.
+					testCtx.Scheduler.SchedulingQueue.Activate(logger, map[string]*v1.Pod{scenario.podGatedInQueue: st.MakePod().Namespace(testCtx.NS.Name).Name(scenario.podGatedInQueue).Obj()})
+					if !podInUnschedulablePodPool(t, testCtx.Scheduler.SchedulingQueue, scenario.podGatedInQueue) {
+						t.Fatalf("Expected the pod %s to be in the queue even after the activation", scenario.podGatedInQueue)
+					}
+				case scenario.podRunningPreemption != nil:
+					if err := wait.PollUntilContextTimeout(testCtx.Ctx, time.Millisecond*200, wait.ForeverTestTimeout, false, func(ctx context.Context) (bool, error) {
+						return preemptionPlugin.Evaluator.IsPodRunningPreemption(createdPods[*scenario.podRunningPreemption].GetUID()), nil
+					}); err != nil {
+						t.Fatalf("Expected the pod %s to be running preemption", createdPods[*scenario.podRunningPreemption].Name)
+					}
+				}
+			}
+		})
+	}
+}
+
+// podInUnschedulablePodPool checks if the given Pod is in the unschedulable pod pool.
+func podInUnschedulablePodPool(t *testing.T, queue queue.SchedulingQueue, podName string) bool {
+	t.Helper()
+	// First, look for the pod in the activeQ.
+	for _, pod := range queue.PodsInActiveQ() {
+		if pod.Name == podName {
+			return false
+		}
+	}
+
+	pendingPods, _ := queue.PendingPods()
+	for _, pod := range pendingPods {
+		if pod.Name == podName {
+			return true
+		}
+	}
+	return false
+}
 
 // TestNonPreemption tests NonPreempt option of PriorityClass of scheduler works as expected.
 func TestNonPreemption(t *testing.T) {
@@ -554,8 +1068,10 @@ func TestNonPreemption(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Error creating nodes: %v", err)
 	}
+
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				defer testutils.CleanupPods(testCtx.Ctx, cs, t, []*v1.Pod{preemptor, victim})
 				preemptor.Spec.PreemptionPolicy = test.PreemptionPolicy
 				victimPod, err := createPausePod(cs, victim)
@@ -582,6 +1098,7 @@ func TestNonPreemption(t *testing.T) {
 			})
 		}
 	}
+}
 
 // TestDisablePreemption tests disable pod preemption of scheduler works as expected.
 func TestDisablePreemption(t *testing.T) {
@@ -630,8 +1147,9 @@ func TestDisablePreemption(t *testing.T) {
 		t.Fatalf("Error creating nodes: %v", err)
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				pods := make([]*v1.Pod, len(test.existingPods))
 				// Create and run existingPods.
 				for i, p := range test.existingPods {
@@ -661,6 +1179,7 @@ func TestDisablePreemption(t *testing.T) {
 			})
 		}
 	}
+}
 
 // This test verifies that system critical priorities are created automatically and resolved properly.
 func TestPodPriorityResolution(t *testing.T) {
@@ -736,9 +1255,9 @@ func TestPodPriorityResolution(t *testing.T) {
 	}
 
 	pods := make([]*v1.Pod, 0, len(tests))
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.Name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.Name, asyncPreemptionEnabled), func(t *testing.T) {
-			t.Run(test.Name, func(t *testing.T) {
 				pod, err := runPausePod(cs, test.Pod)
 				if err != nil {
 					if test.ExpectedError == nil {
@@ -757,10 +1276,10 @@ func TestPodPriorityResolution(t *testing.T) {
 				} else {
 					t.Errorf("Expected pod %v to have priority %v but was nil", pod.Name, test.PriorityClass)
 				}
-			})
+				testutils.CleanupPods(testCtx.Ctx, cs, t, pods)
 			})
 		}
-	testutils.CleanupPods(testCtx.Ctx, cs, t, pods)
+	}
 	testutils.CleanupNodes(cs, t)
 }
 
@@ -824,8 +1343,9 @@ func TestPreemptionStarvation(t *testing.T) {
 		t.Fatalf("Error creating nodes: %v", err)
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				pendingPods := make([]*v1.Pod, test.numExpectedPending)
 				numRunningPods := test.numExistingPod - test.numExpectedPending
 				runningPods := make([]*v1.Pod, numRunningPods)
@@ -878,6 +1398,7 @@ func TestPreemptionStarvation(t *testing.T) {
 			})
 		}
 	}
+}
 
 // TestPreemptionRaces tests that other scheduling events and operations do not
 // race with the preemption process.
@@ -924,8 +1445,9 @@ func TestPreemptionRaces(t *testing.T) {
 		t.Fatalf("Error creating nodes: %v", err)
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				if test.numRepetitions <= 0 {
 					test.numRepetitions = 1
 				}
@@ -992,6 +1514,7 @@ func TestPreemptionRaces(t *testing.T) {
 			})
 		}
 	}
+}
 
 const (
 	alwaysFailPlugin = "alwaysFailPlugin"
@@ -1126,8 +1649,9 @@ func TestNominatedNodeCleanUp(t *testing.T) {
 		},
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", tt.name, asyncPreemptionEnabled), func(t *testing.T) {
 				cfg := configtesting.V1ToInternalWithDefaults(t, configv1.KubeSchedulerConfiguration{
 					Profiles: []configv1.KubeSchedulerProfile{{
 						SchedulerName: pointer.String(v1.DefaultSchedulerName),
@@ -1197,6 +1721,7 @@ func TestNominatedNodeCleanUp(t *testing.T) {
 			})
 		}
 	}
+}
 
 func mkMinAvailablePDB(name, namespace string, uid types.UID, minAvailable int, matchLabels map[string]string) *policy.PodDisruptionBudget {
 	intMinAvailable := intstr.FromInt32(int32(minAvailable))
@@ -1405,8 +1930,9 @@ func TestPDBInPreemption(t *testing.T) {
 		},
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				for i := 1; i <= test.nodeCnt; i++ {
 					nodeName := fmt.Sprintf("node-%v", i)
 					_, err := createNode(cs, st.MakeNode().Name(nodeName).Capacity(defaultNodeRes).Obj())
@@ -1482,6 +2008,7 @@ func TestPDBInPreemption(t *testing.T) {
 			})
 		}
 	}
+}
 
 func initTestPreferNominatedNode(t *testing.T, nsPrefix string, opts ...scheduler.Option) *testutils.TestContext {
 	testCtx := testutils.InitTestSchedulerWithOptions(t, testutils.InitTestAPIServer(t, nsPrefix, nil), 0, opts...)
@@ -1563,8 +2090,9 @@ func TestPreferNominatedNode(t *testing.T) {
 		},
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				testCtx := initTestPreferNominatedNode(t, "perfer-nominated-node")
 				cs := testCtx.ClientSet
 				nsName := testCtx.NS.Name
@@ -1611,6 +2139,7 @@ func TestPreferNominatedNode(t *testing.T) {
 			})
 		}
 	}
+}
 
 // TestReadWriteOncePodPreemption tests preemption scenarios for pods with
 // ReadWriteOncePod PVCs.
@@ -1912,8 +2441,9 @@ func TestReadWriteOncePodPreemption(t *testing.T) {
 		t.Fatalf("Error creating node: %v", err)
 	}
 
+	for _, asyncPreemptionEnabled := range []bool{true, false} {
 		for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
+			t.Run(fmt.Sprintf("%s (Async preemption enabled: %v)", test.name, asyncPreemptionEnabled), func(t *testing.T) {
 				if err := test.init(); err != nil {
 					t.Fatalf("Error while initializing test: %v", err)
 				}
@@ -1961,3 +2491,4 @@ func TestReadWriteOncePodPreemption(t *testing.T) {
 			})
 		}
 	}
+}

test/integration/scheduler_perf/misc/performance-config.yaml

@@ -354,10 +354,20 @@
       initPods: 2000
       measurePods: 500
   - name: 5000Nodes
-    featureGates:
-      SchedulerQueueingHints: false
     labels: [performance]
     threshold: 200
+    featureGates:
+      SchedulerQueueingHints: false
+      SchedulerAsyncPreemption: false
+    params:
+      initNodes: 5000
+      initPods: 20000
+      measurePods: 5000
+  - name: 5000Nodes_AsyncPreemptionEnabled
+    threshold: 200
+    labels: [performance]
+    featureGates:
+      SchedulerAsyncPreemption: true
     params:
       initNodes: 5000
       initPods: 20000

test/integration/scheduler_perf/templates/pod-high-priority.yaml

@@ -1,7 +1,7 @@
 apiVersion: v1
 kind: Pod
 metadata:
-  generateName: pod-
+  generateName: pod-high-priority-
 spec:
   priority: 10
   containers: