[cherry-pick for release-1.8]add ignored csi provisioner when compute csi resources

cmd/scheduler/app/options/options.go

@@ -79,6 +79,11 @@ type ServerOption struct {
 	NodeSelector      []string
 	EnableCacheDumper bool
 	NodeWorkerThreads uint32
+
+	// IgnoredCSIProvisioners contains a list of provisioners, and pod request pvc with these provisioners will
+	// not be counted in pod pvc resource request and node.Allocatable, because the spec.drivers of csinode resource
+	// is always null, these provisioners usually are host path csi controllers like rancher.io/local-path and hostpath.csi.k8s.io.
+	IgnoredCSIProvisioners []string
 }
 
 type DecryptFunc func(c *ServerOption) error
@@ -134,6 +139,7 @@ func (s *ServerOption) AddFlags(fs *pflag.FlagSet) {
 	fs.StringSliceVar(&s.NodeSelector, "node-selector", nil, "volcano only work with the labeled node, like: --node-selector=volcano.sh/role:train --node-selector=volcano.sh/role:serving")
 	fs.BoolVar(&s.EnableCacheDumper, "cache-dumper", true, "Enable the cache dumper, it's true by default")
 	fs.Uint32Var(&s.NodeWorkerThreads, "node-worker-threads", defaultNodeWorkers, "The number of threads syncing node operations.")
+	fs.StringSliceVar(&s.IgnoredCSIProvisioners, "ignored-provisioners", nil, "The provisioners that will be ignored during pod pvc request computation and preemption.")
 }
 
 // CheckOptionOrDie check lock-object-namespace when LeaderElection is enabled.

pkg/scheduler/cache/cache.go

@@ -74,6 +74,9 @@ const (
 	defaultMetricsInternal = 30 * time.Second
 )
 
+// defaultIgnoredProvisioners contains provisioners that will be ignored during pod pvc request computation and preemption.
+var defaultIgnoredProvisioners = []string{"rancher.io/local-path", "hostpath.csi.k8s.io"}
+
 func init() {
 	schemeBuilder := runtime.SchemeBuilder{
 		v1.AddToScheme,
@@ -83,8 +86,8 @@ func init() {
 }
 
 // New returns a Cache implementation.
-func New(config *rest.Config, schedulerNames []string, defaultQueue string, nodeSelectors []string, nodeWorkers uint32) Cache {
-	return newSchedulerCache(config, schedulerNames, defaultQueue, nodeSelectors, nodeWorkers)
+func New(config *rest.Config, schedulerNames []string, defaultQueue string, nodeSelectors []string, nodeWorkers uint32, ignoredProvisioners []string) Cache {
+	return newSchedulerCache(config, schedulerNames, defaultQueue, nodeSelectors, nodeWorkers, ignoredProvisioners)
 }
 
 // SchedulerCache cache for the kube batch
@@ -148,6 +151,11 @@ type SchedulerCache struct {
 	imageStates map[string]*imageState
 
 	nodeWorkers uint32
+
+	// IgnoredCSIProvisioners contains a list of provisioners, and pod request pvc with these provisioners will
+	// not be counted in pod pvc resource request and node.Allocatable, because the spec.drivers of csinode resource
+	// is always null, these provisioners usually are host path csi controllers like rancher.io/local-path and hostpath.csi.k8s.io.
+	IgnoredCSIProvisioners sets.Set[string]
 }
 
 type imageState struct {
@@ -390,7 +398,7 @@ func (pgb *podgroupBinder) Bind(job *schedulingapi.JobInfo, cluster string) (*sc
 	return job, nil
 }
 
-func newSchedulerCache(config *rest.Config, schedulerNames []string, defaultQueue string, nodeSelectors []string, nodeWorkers uint32) *SchedulerCache {
+func newSchedulerCache(config *rest.Config, schedulerNames []string, defaultQueue string, nodeSelectors []string, nodeWorkers uint32, ignoredProvisioners []string) *SchedulerCache {
 	kubeClient, err := kubernetes.NewForConfig(config)
 	if err != nil {
 		panic(fmt.Sprintf("failed init kubeClient, with err: %v", err))
@@ -453,6 +461,13 @@ func newSchedulerCache(config *rest.Config, schedulerNames []string, defaultQueu
 		NodeList:    []string{},
 		nodeWorkers: nodeWorkers,
 	}
+
+	ignoredProvisionersSet := sets.New[string]()
+	for _, provisioner := range append(ignoredProvisioners, defaultIgnoredProvisioners...) {
+		ignoredProvisionersSet.Insert(provisioner)
+	}
+	sc.IgnoredCSIProvisioners = ignoredProvisionersSet
+
 	if len(nodeSelectors) > 0 {
 		for _, nodeSelectorLabel := range nodeSelectors {
 			nodeSelectorLabelLen := len(nodeSelectorLabel)

pkg/scheduler/cache/event_handlers.go

@@ -127,7 +127,12 @@ func (sc *SchedulerCache) getPodCSIVolumes(pod *v1.Pod) (map[v1.ResourceName]int
 				return volumes, err
 			}
 		}
+
 		driverName := sc.getCSIDriverInfo(pvc)
+		if sc.isIgnoredProvisioner(driverName) {
+			klog.V(5).InfoS("Provisioner ignored, skip count pod pvc num", "driverName", driverName)
+			continue
+		}
 		if driverName == "" {
 			klog.V(5).InfoS("Could not find a CSI driver name for pvc(%s/%s), not counting volume", pvc.Namespace, pvc.Name)
 			continue
@@ -147,6 +152,10 @@ func (sc *SchedulerCache) getPodCSIVolumes(pod *v1.Pod) (map[v1.ResourceName]int
 	return volumes, nil
 }
 
+func (sc *SchedulerCache) isIgnoredProvisioner(driverName string) bool {
+	return sc.IgnoredCSIProvisioners.Has(driverName)
+}
+
 func (sc *SchedulerCache) getCSIDriverInfo(pvc *v1.PersistentVolumeClaim) string {
 	pvName := pvc.Spec.VolumeName
 

pkg/scheduler/plugins/predicates/predicates_test.go

@@ -70,7 +70,7 @@ func TestEventHandler(t *testing.T) {
 		return
 	}
 
-	sc := cache.New(config, option.SchedulerNames, option.DefaultQueue, option.NodeSelector, option.NodeWorkerThreads)
+	sc := cache.New(config, option.SchedulerNames, option.DefaultQueue, option.NodeSelector, option.NodeWorkerThreads, nil)
 	schedulerCache := sc.(*cache.SchedulerCache)
 
 	// pending pods

pkg/scheduler/scheduler.go

@@ -64,7 +64,7 @@ func NewScheduler(config *rest.Config, opt *options.ServerOption) (*Scheduler, e
 		}
 	}
 
-	cache := schedcache.New(config, opt.SchedulerNames, opt.DefaultQueue, opt.NodeSelector, opt.NodeWorkerThreads)
+	cache := schedcache.New(config, opt.SchedulerNames, opt.DefaultQueue, opt.NodeSelector, opt.NodeWorkerThreads, opt.IgnoredCSIProvisioners)
 	scheduler := &Scheduler{
 		schedulerConf:  opt.SchedulerConf,
 		fileWatcher:    watcher,