feat: refactor device affinity and support strict device affinity requirement

Author: JustinChengLZ

cmd/katalyst-agent/app/options/qrm/gpu_plugin.go

@@ -30,6 +30,7 @@ type GPUOptions struct {
 	GPUMemoryAllocatablePerGPU string
 	SkipGPUStateCorruption     bool
 	RDMADeviceNames            []string
+	RequiredDeviceAffinity     bool
 
 	GPUStrategyOptions *gpustrategy.GPUStrategyOptions
 }
@@ -55,6 +56,8 @@ func (o *GPUOptions) AddFlags(fss *cliflag.NamedFlagSets) {
 	fs.BoolVar(&o.SkipGPUStateCorruption, "skip-gpu-state-corruption",
 		o.SkipGPUStateCorruption, "skip gpu state corruption, and it will be used after updating state properties")
 	fs.StringSliceVar(&o.RDMADeviceNames, "rdma-resource-names", o.RDMADeviceNames, "The name of the RDMA resource")
+	fs.BoolVar(&o.RequiredDeviceAffinity, "gpu-required-device-affinity", o.RequiredDeviceAffinity,
+		"required device affinity, and when true it will cause pods to admit fail if unable to meet device affinity")
 	o.GPUStrategyOptions.AddFlags(fss)
 }
 
@@ -71,5 +74,6 @@ func (o *GPUOptions) ApplyTo(conf *qrmconfig.GPUQRMPluginConfig) error {
 	if err := o.GPUStrategyOptions.ApplyTo(conf.GPUStrategyConfig); err != nil {
 		return err
 	}
+	conf.RequiredDeviceAffinity = o.RequiredDeviceAffinity
 	return nil
 }

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

@@ -246,7 +246,7 @@ func (p *BasePlugin) UpdateAllocatableAssociatedDevices(
 		deviceTopology.Devices[device.ID] = machine.DeviceInfo{
 			Health:         device.Health,
 			NumaNodes:      numaNode,
-			DeviceAffinity: make(map[machine.AffinityPriority]machine.DeviceIDs),
+			DeviceAffinity: make(map[machine.Dimension]machine.DeviceIDs),
 		}
 	}