fix: determine if the default resource state should be init at start

Author: JustinChengLZ

pkg/agent/qrm-plugins/gpu/baseplugin/base.go

@@ -291,15 +291,19 @@ func (p *BasePlugin) GenerateMachineStateFromPodEntries(
 // For example, we may have multiple device names for a same device type, e.g. "nvidia.com/gpu" and "hw.com/npu",
 // so we map them to the same device type, which allows us to allocate them interchangeably.
 func (p *BasePlugin) RegisterDeviceNames(deviceNames []string, deviceType string) {
-	for _, deviceeName := range deviceNames {
-		p.deviceNameToTypeMap[deviceeName] = deviceType
+	for _, deviceName := range deviceNames {
+		p.deviceNameToTypeMap[deviceName] = deviceType
 		if _, ok := p.deviceTypeToNames[deviceType]; !ok {
 			p.deviceTypeToNames[deviceType] = sets.NewString()
 		}
-		p.deviceTypeToNames[deviceType].Insert(deviceeName)
+		p.deviceTypeToNames[deviceType].Insert(deviceName)
 	}
 }
 
+func (p *BasePlugin) GetDeviceNameToTypeMap() map[string]string {
+	return p.deviceNameToTypeMap
+}
+
 // ResolveResourceName takes in a resourceName and tries to find a mapping of resource type from deviceNameToTypeMap.
 // If no mapping is found, resourceName is returned if fallback is true. If fallback is false, an empty string is returned.
 func (p *BasePlugin) ResolveResourceName(resourceName string, fallback bool) string {

pkg/agent/qrm-plugins/gpu/baseplugin/reporter/reporter.go

@@ -206,9 +206,9 @@ func (p *gpuReporterPlugin) GetReportContent(ctx context.Context, _ *v1alpha1.Em
 
 func (p *gpuReporterPlugin) buildReportResponse() (*v1alpha1.GetReportContentResponse, error) {
 	// The reporter picks the latest topology from all configured GPU devices to report to CNR.
-	topologiesMap, err := p.deviceTopologyRegistry.GetDeviceTopologies(p.gpuDeviceNames)
-	if err != nil {
-		return nil, err
+	topologiesMap, ok := p.deviceTopologyRegistry.GetDeviceTopologies(p.gpuDeviceNames)
+	if !ok {
+		return nil, fmt.Errorf("failed to get any device topology")
 	}
 	latestDeviceTopology := machine.PickLatestDeviceTopology(topologiesMap)
 
@@ -273,7 +273,7 @@ func (p *gpuReporterPlugin) getTopologyZoneReportField(topologiesMap map[string]
 		return nil, fmt.Errorf("no zone resources found for device topology")
 	}
 
-	zoneAllocations := p.getZoneAllocations(machineState)
+	zoneAllocations := p.getZoneAllocations(topologiesMap, machineState)
 
 	generatedTopologyZones := topologyZoneGenerator.GenerateTopologyZoneStatus(zoneAllocations, zoneResources,
 		zoneAttributes, nil, nil, nil)
@@ -424,7 +424,7 @@ func (p *gpuReporterPlugin) getZoneResources(topologiesMap map[string]*machine.D
 }
 
 // getZoneAllocations returns the map of gpu zone nodes to their pod allocations
-func (p *gpuReporterPlugin) getZoneAllocations(machineState state.AllocationResourcesMap) map[util.ZoneNode]util.ZoneAllocations {
+func (p *gpuReporterPlugin) getZoneAllocations(topologiesMap map[string]*machine.DeviceTopology, machineState state.AllocationResourcesMap) map[util.ZoneNode]util.ZoneAllocations {
 	// First construct map of device id to allocations
 	idToAllocations := make(map[string]util.ZoneAllocations)
 
@@ -444,6 +444,11 @@ func (p *gpuReporterPlugin) getZoneAllocations(machineState state.AllocationReso
 
 					// Override the resource name if there is a specified device name
 					if allocInfo.DeviceName != "" {
+						// Skip reporting if it is not a GPU device
+						if _, ok := topologiesMap[allocInfo.DeviceName]; !ok {
+							continue
+						}
+
 						resourceName = v1.ResourceName(allocInfo.DeviceName)
 					}
 

pkg/agent/qrm-plugins/gpu/baseplugin/reporter/reporter_test.go

@@ -2216,6 +2216,61 @@ func TestGpuReporterPlugin_GetReportContent(t *testing.T) {
 				},
 			},
 		},
+		{
+			name: "Allocations are skipped when allocation DeviceName is not a GPU device",
+			deviceTopology: &machine.DeviceTopology{
+				PriorityDimensions: []string{"numa"},
+				Devices: map[string]machine.DeviceInfo{
+					"gpu-0": {
+						Health:         pluginapi.Healthy,
+						NumaNodes:      []int{0},
+						DeviceAffinity: map[machine.AffinityPriority]machine.DeviceIDs{{PriorityLevel: 0, Dimension: machine.Dimension{Name: "numa", Value: "0"}}: {}},
+					},
+				},
+			},
+			machineTopology: []cadvisorapi.Node{{Id: 0, Cores: []cadvisorapi.Core{{SocketID: 0, Id: 0, Threads: []int{0, 4}}}}},
+			machineState: state.AllocationResourcesMap{
+				v1.ResourceName("test-gpu-resource"): state.AllocationMap{
+					"gpu-0": {
+						Allocatable: 1,
+						PodEntries: state.PodEntries{
+							"pod-uid-0": state.ContainerEntries{
+								"c0": &state.AllocationInfo{
+									AllocationMeta:      commonstate.AllocationMeta{PodUid: "pod-uid-0", PodNamespace: "default", PodName: "p0", ContainerName: "c0"},
+									DeviceName:          "not-a-gpu-device",
+									AllocatedAllocation: state.Allocation{Quantity: 1, NUMANodes: []int{0}},
+								},
+							},
+						},
+					},
+				},
+			},
+			expectedSpec: []*nodev1alpha1.Property{{PropertyName: propertyNameGPUTopology, PropertyValues: []string{"numa"}}},
+			expectedStatus: []*nodev1alpha1.TopologyZone{
+				{
+					Type: nodev1alpha1.TopologyTypeSocket,
+					Name: "0",
+					Children: []*nodev1alpha1.TopologyZone{
+						{
+							Type: nodev1alpha1.TopologyTypeNuma,
+							Name: "0",
+							Children: []*nodev1alpha1.TopologyZone{
+								{
+									Type:       nodev1alpha1.TopologyTypeGPU,
+									Name:       "gpu-0",
+									Attributes: []nodev1alpha1.Attribute{{Name: "numa", Value: "0"}},
+									Resources: nodev1alpha1.Resources{
+										Allocatable: &v1.ResourceList{"test-gpu": resource.MustParse("1")},
+										Capacity:    &v1.ResourceList{"test-gpu": resource.MustParse("1")},
+									},
+									Allocations: []*nodev1alpha1.Allocation{},
+								},
+							},
+						},
+					},
+				},
+			},
+		},
 		{
 			name:           "Merge resources across multiple device names for same device type",
 			gpuDeviceNames: []string{"test-gpu-a", "test-gpu-b"},

pkg/agent/qrm-plugins/gpu/customdeviceplugin/gpu/gpu.go

@@ -55,7 +55,7 @@ func NewGPUDevicePlugin(base *baseplugin.BasePlugin) customdeviceplugin.CustomDe
 	// GPUDeviceType is the key used for GPU state management in the QRM framework,
 	// while GPUDeviceNames are the actual resource names used to fetch the GPU device topologies.
 	base.DefaultResourceStateGeneratorRegistry.RegisterResourceStateGenerator(gpuconsts.GPUDeviceType,
-		state.NewGenericDefaultResourceStateGenerator(base.Conf.GPUDeviceNames, base.DeviceTopologyRegistry, 1))
+		state.NewGenericDefaultResourceStateGenerator(base.Conf.GPUDeviceNames, base.DeviceTopologyRegistry, 1, true))
 	base.RegisterDeviceNames(base.Conf.GPUDeviceNames, gpuconsts.GPUDeviceType)
 
 	return &GPUDevicePlugin{
@@ -149,6 +149,7 @@ func (p *GPUDevicePlugin) AllocateAssociatedDevice(
 			qosLevel,
 			deviceReq.DeviceName,
 			"",
+			p.GetDeviceNameToTypeMap(),
 		)
 		if err != nil {
 			return nil, fmt.Errorf("GPU allocation using strategy failed: %v", err)

pkg/agent/qrm-plugins/gpu/customdeviceplugin/rdma/rdma.go

@@ -49,8 +49,8 @@ func NewRDMADevicePlugin(base *baseplugin.BasePlugin) customdeviceplugin.CustomD
 	// RDMADeviceType is the key used for RDMA state management in the QRM framework,
 	// while RDMADeviceNames are the actual resource names used to fetch the RDMA device topologies
 	base.DefaultResourceStateGeneratorRegistry.RegisterResourceStateGenerator(gpuconsts.RDMADeviceType,
-		state.NewGenericDefaultResourceStateGenerator(base.Conf.RDMADeviceNames, base.DeviceTopologyRegistry))
-	base.RegisterDeviceNameToType(base.Conf.RDMADeviceNames, gpuconsts.RDMADeviceType)
+		state.NewGenericDefaultResourceStateGenerator(base.Conf.RDMADeviceNames, base.DeviceTopologyRegistry, 1, false))
+	base.RegisterDeviceNames(base.Conf.RDMADeviceNames, gpuconsts.RDMADeviceType)
 
 	return &RDMADevicePlugin{
 		BasePlugin:  base,
@@ -139,6 +139,7 @@ func (p *RDMADevicePlugin) AllocateAssociatedDevice(
 		qosLevel,
 		rdmaDeviceName,
 		accompanyResourceName,
+		p.GetDeviceNameToTypeMap(),
 	)
 	if err != nil {
 		return nil, fmt.Errorf("RDMA allocation using strategy failed: %v", err)
@@ -170,6 +171,7 @@ func (p *RDMADevicePlugin) AllocateAssociatedDevice(
 			Quantity:  float64(len(allocatedRdmaDevices)),
 			NUMANodes: hintNodes.ToSliceInt(),
 		},
+		DeviceName: deviceReq.DeviceName,
 	}
 
 	allocationInfo.TopologyAwareAllocations = topologyAwareAllocations

pkg/agent/qrm-plugins/gpu/customdeviceplugin/rdma/rdma_test.go

@@ -32,6 +32,7 @@ import (
 	"github.com/kubewharf/katalyst-core/pkg/agent/qrm-plugins/gpu/baseplugin"
 	gpuconsts "github.com/kubewharf/katalyst-core/pkg/agent/qrm-plugins/gpu/consts"
 	"github.com/kubewharf/katalyst-core/pkg/agent/qrm-plugins/gpu/state"
+	gpuutil "github.com/kubewharf/katalyst-core/pkg/agent/qrm-plugins/gpu/util"
 	"github.com/kubewharf/katalyst-core/pkg/config"
 	"github.com/kubewharf/katalyst-core/pkg/config/agent/qrm/statedirectory"
 	"github.com/kubewharf/katalyst-core/pkg/metaserver"
@@ -100,6 +101,9 @@ func makeTestBasePlugin(t *testing.T) *baseplugin.BasePlugin {
 	conf.StateDirectoryConfiguration = &statedirectory.StateDirectoryConfiguration{
 		StateFileDirectory: tmpDir,
 	}
+	// Ensure both GPU and RDMA device names are configured so that
+	// gpuutil.ResolveResourceName can resolve logical types in tests
+	conf.GPUDeviceNames = []string{"test-gpu"}
 	conf.RDMADeviceNames = []string{"test-rdma"}
 
 	basePlugin, err := baseplugin.NewBasePlugin(agentCtx, conf, metrics.DummyMetrics{})
@@ -110,13 +114,14 @@ func makeTestBasePlugin(t *testing.T) *baseplugin.BasePlugin {
 
 	basePlugin.SetState(stateImpl)
 
-	// Register gpu device type and gpu device topology provider as it is an accompany resource for rdma
-	basePlugin.RegisterDeviceNameToType([]string{"test-gpu"}, gpuconsts.GPUDeviceType)
 	gpuTopologyProvider := machine.NewDeviceTopologyProvider()
 	basePlugin.DeviceTopologyRegistry.RegisterDeviceTopologyProvider(gpuconsts.GPUDeviceType, gpuTopologyProvider)
 	// Also register a provider for rdma device type to align with plugin lookup by type
 	basePlugin.DeviceTopologyRegistry.RegisterDeviceTopologyProvider(gpuconsts.RDMADeviceType, machine.NewDeviceTopologyProvider())
 
+	// Register GPU device names for resource name resolution in tests.
+	basePlugin.RegisterDeviceNames(conf.GPUDeviceNames, gpuconsts.GPUDeviceType)
+
 	return basePlugin
 }
 
@@ -466,7 +471,7 @@ func TestRDMADevicePlugin_AllocateAssociatedDevices(t *testing.T) {
 			}
 
 			if tt.accompanyResourceAllocationInfo != nil && tt.accompanyResourceName != "" {
-				accompanyResourceType := basePlugin.ResolveResourceName(tt.accompanyResourceName, false)
+				accompanyResourceType := gpuutil.ResolveResourceName(basePlugin.GetDeviceNameToTypeMap(), tt.accompanyResourceName, false)
 				assert.NotEmpty(t, accompanyResourceType)
 				basePlugin.GetState().SetAllocationInfo(v1.ResourceName(accompanyResourceType), tt.podUID, tt.containerName, tt.accompanyResourceAllocationInfo, false)
 			}
@@ -480,7 +485,7 @@ func TestRDMADevicePlugin_AllocateAssociatedDevices(t *testing.T) {
 			}
 
 			if tt.accompanyResourceName != "" && tt.accompanyDeviceTopology != nil {
-				accompanyResourceType := basePlugin.ResolveResourceName(tt.accompanyResourceName, false)
+				accompanyResourceType := gpuutil.ResolveResourceName(basePlugin.GetDeviceNameToTypeMap(), tt.accompanyResourceName, false)
 				assert.NotEmpty(t, accompanyResourceType)
 				err := basePlugin.DeviceTopologyRegistry.SetDeviceTopology(accompanyResourceType, tt.accompanyDeviceTopology)
 				assert.NoError(t, err)

pkg/agent/qrm-plugins/gpu/resourceplugin/gpumemory/gpu_mem.go

@@ -52,7 +52,8 @@ func NewGPUMemPlugin(base *baseplugin.BasePlugin) resourceplugin.ResourcePlugin
 	// string(consts.ResourceGPUMemory) is the key used for state management in the QRM framework,
 	// while GPUDeviceNames are the actual resource names used to fetch the device topologies.
 	base.DefaultResourceStateGeneratorRegistry.RegisterResourceStateGenerator(string(consts.ResourceGPUMemory),
-		state.NewGenericDefaultResourceStateGenerator(base.Conf.GPUDeviceNames, base.DeviceTopologyRegistry, float64(base.Conf.GPUMemoryAllocatablePerGPU.Value())))
+		state.NewGenericDefaultResourceStateGenerator(base.Conf.GPUDeviceNames, base.DeviceTopologyRegistry,
+			float64(base.Conf.GPUMemoryAllocatablePerGPU.Value()), true))
 	return &GPUMemPlugin{
 		BasePlugin: base,
 	}
@@ -535,6 +536,7 @@ func (p *GPUMemPlugin) Allocate(
 		qosLevel,
 		deviceReq.DeviceName,
 		"",
+		p.GetDeviceNameToTypeMap(),
 	)
 	if err != nil {
 		return nil, fmt.Errorf("GPU allocation using strategy failed: %v", err)

pkg/agent/qrm-plugins/gpu/state/generator_stub.go

@@ -24,6 +24,10 @@ func NewDefaultResourceStateGeneratorStub() DefaultResourceStateGenerator {
 	return &defaultResourceStateGeneratorStub{}
 }
 
+func (d *defaultResourceStateGeneratorStub) MustInitDefaultResourceState() bool {
+	return true
+}
+
 func (d *defaultResourceStateGeneratorStub) GenerateDefaultResourceState() (AllocationMap, error) {
 	return map[string]*AllocationState{
 		"gpu-0": {

pkg/agent/qrm-plugins/gpu/state/interface.go

@@ -49,6 +49,9 @@ type writer interface {
 // DefaultResourceStateGenerator interface is used to generate default resource state for each resource
 type DefaultResourceStateGenerator interface {
 	GenerateDefaultResourceState() (AllocationMap, error)
+
+	// MustInitDefaultResourceState indicates whether the default resource state of a resource should be initialized at the start
+	MustInitDefaultResourceState() bool
 }
 
 // ReadonlyState interface only provides methods for tracking pod assignments

pkg/agent/qrm-plugins/gpu/state/state.go

@@ -32,7 +32,9 @@ type AllocationInfo struct {
 
 	AllocatedAllocation      Allocation            `json:"allocated_allocation"`
 	TopologyAwareAllocations map[string]Allocation `json:"topology_aware_allocations"`
-	DeviceName               string                `json:"device_name"`
+	// DeviceName will be empty if it is a resource allocation (e.g. gpu memory), but it will be non-empty if
+	// it is a device allocation (e.g. gpu, rdma)
+	DeviceName string `json:"device_name"`
 }
 
 type Allocation struct {