Enhance the scheduling performance of karmada-schduler

[LivingCcj]

When karmada-scheduler schedules the rb object for the first time, it would respectively calculate the number of available replicas in the subcluster in the Select phase and the AssignReplica phase. The picture below records metric of karmada_scheduler_scheduling_algorithm_duration_seconds_bucket.

there could be optimized to once calculate, The result calculated in the Select phase could be directly repeated in the AssignReplica phase. In order to improve the qps of the karmada-scheduler
[](

karmada/pkg/scheduler/core/generic_scheduler.go

Lines 97 to 107 in d4c5890

	clusters, err := g.selectClusters(clustersScore, spec.Placement, spec)
	if err != nil {
	return result, fmt.Errorf("failed to select clusters: %w", err)
	}
	klog.V(4).Infof("Selected clusters: %v", clusters)
	clustersWithReplicas, err := g.assignReplicas(clusters, spec, status)
	if err != nil {
	return result, fmt.Errorf("failed to assign replicas: %w", err)
	}
	klog.V(4).Infof("Assigned Replicas: %v", clustersWithReplicas)

)

[chaosi-zju]

Since the time interval between the selectClusters step and assignReplicas step is only 20 milliseconds, it does seem unnecessary to calculate availableReplicas twice. So, I also think that the second calculation can be omitted.

Hi @Garrybest ,what is your opinion?

[LivingCcj]

In the scheduler step, there have 4 steps, and the findClustersThatFit & priortizeClusters only cost less time, the selectClusters or assignReplicas would cost more time from the picture.

The large cost time will be happened in calculate the available replicas for each member cluster. and the selectClusters and assignReplicas step will invoke the calAvailableReplicas twice in followed picture.

There is an optimization points：
In the selectClusters step, that has got the available replicas for each member cluster, and the result could be passed to the next assignReplicas step. it could reduce to invoke the calAvailableReplicas function for once.

[RainbowMango]

cc @zhzhuang-zju

[LivingCcj]

first，update some structs

ClusterDetailInfo struct add AvailableClusterReplicas field, which indicate the cluster and the available replicas.

// ClusterDetailInfo indicate the cluster information
type ClusterDetailInfo struct {
	Name              string
	Score             int64
	AvailableReplicas int64

	Cluster                  *clusterv1alpha1.Cluster
	AvailableClusterReplicas AvailableClusterReplicas // new field 
}

// AvailableClusterReplicas is used to record the latest available replicas of the cluster.
type AvailableClusterReplicas struct {
	Cluster  *clusterv1alpha1.Cluster
	Replicas int64
}

assignState struct add availableClusterReplicas field, which indicate the cluster and the available replicas from select cluster step.

// assignState is a wrapper of the input for assigning function.
type assignState struct {
	candidates []*clusterv1alpha1.Cluster
	strategy   *policyv1alpha1.ReplicaSchedulingStrategy
	spec       *workv1alpha2.ResourceBindingSpec

	// fields below are indirect results
	strategyType string

	// assignmentMode represents the mode how to assign replicas
	assignmentMode assignmentMode

	scheduledClusters []workv1alpha2.TargetCluster
	assignedReplicas  int32
	availableClusters []workv1alpha2.TargetCluster
	availableReplicas int32

	// targetReplicas is the replicas that we need to schedule in this round 
	targetReplicas int32

	//availableClusterReplicas is the available replicas of candidate clusters
	availableClusterReplicas []spreadconstraint.AvailableClusterReplicas // new field 
}

second,refresh the input and output function, and some core functions like

karmada/pkg/scheduler/core/generic_scheduler.go

Lines 97 to 108 in efca312

	clusters, err := g.selectClusters(clustersScore, spec.Placement, spec)
	if err != nil {
	return result, fmt.Errorf("failed to select clusters: %w", err)
	}
	klog.V(4).Infof("Selected clusters: %v", clusters)
	clustersWithReplicas, err := g.assignReplicas(clusters, spec, status)
	if err != nil {
	return result, fmt.Errorf("failed to assign replicas: %w", err)
	}
	klog.V(4).Infof("Assigned Replicas: %v", clustersWithReplicas)

Update the core function
selectClusters(...) ([]*clusterv1alpha1.Cluster, error) to selectClusters(...) ([]spreadconstraint.AvailableClusterReplicas, error)

And update the core function assignReplicas(clusters []*clusterv1alpha1.Cluster,...) ([]workv1alpha2.TargetCluster, error) to assignReplicas(clusters []spreadconstraint.AvailableClusterReplicas, ...) ([]workv1alpha2.TargetCluster, error)

[RainbowMango]

As we talked at the community meeting, I will look at it and you can continue to send a PR for this.

[LivingCcj]

As we talked at the community meeting, I will look at it and you can continue to send a PR for this.

Ok