util

package
v0.4.1 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Apr 21, 2025 License: Apache-2.0 Imports: 18 Imported by: 0

Documentation

Index

Constants

View Source 
const (
	KubeletConfigPath = "/var/lib/kubelet/config.yaml"

	CGroupBasePath   = "/sys/fs/cgroup"
	CGroupDevicePath = CGroupBasePath + "/devices"

	SYSTEMD  CGroupDriver = "systemd"
	CGROUPFS CGroupDriver = "cgroupfs"
)
View Source 
const (
	DomainPrefix = "nvidia.com"

	// VGPUComputePolicyAnnotation none / balance / fixed(default)
	VGPUComputePolicyAnnotation = DomainPrefix + "/vgpu-compute-policy"

	NodeNvidiaDriverVersionLabel = DomainPrefix + "/node-driver-version"
	NodeNvidiaCudaVersionLabel   = DomainPrefix + "/node-cuda-version"

	VGPUNumberResourceName = DomainPrefix + "/vgpu-number"
	VGPUMemoryResourceName = DomainPrefix + "/vgpu-memory"
	VGPUCoreResourceName   = DomainPrefix + "/vgpu-cores"

	// NodeDeviceHeartbeatAnnotation Node device heartbeat time
	NodeDeviceHeartbeatAnnotation = DomainPrefix + "/node-device-heartbeat"
	NodeDeviceRegisterAnnotation  = DomainPrefix + "/node-device-register"
	NodeDeviceTopologyAnnotation  = DomainPrefix + "/node-device-topology"
	DeviceMemoryFactorAnnotation  = DomainPrefix + "/device-memory-factor"

	// PodIncludeGpuTypeAnnotation Specify the GPU type to be used
	PodIncludeGpuTypeAnnotation = DomainPrefix + "/include-gpu-type"
	// PodExcludeGpuTypeAnnotation Specify the GPU type to exclude
	PodExcludeGpuTypeAnnotation = DomainPrefix + "/exclude-gpu-type"

	// Scheduling strategies at the node and device levels
	NodeSchedulerPolicyAnnotation   = DomainPrefix + "/node-scheduler-policy"
	DeviceSchedulerPolicyAnnotation = DomainPrefix + "/device-scheduler-policy"

	// DeviceTopologyModeAnnotation Specify device topology mode
	DeviceTopologyModeAnnotation = DomainPrefix + "/device-topology-mode"

	// PodIncludeGPUUUIDAnnotation Specify the GPU UUID to be used
	PodIncludeGPUUUIDAnnotation = DomainPrefix + "/include-gpu-uuid"
	// PodExcludeGPUUUIDAnnotation Specify the GPU UUID to be excluded
	PodExcludeGPUUUIDAnnotation = DomainPrefix + "/exclude-gpu-uuid"

	PodPredicateNodeAnnotation = DomainPrefix + "/predicate-node"
	PodPredicateTimeAnnotation = DomainPrefix + "/predicate-time"
	PodAssignedPhaseLabel      = DomainPrefix + "/assigned-phase"

	// PodVGPUPreAllocAnnotation Pre allocated device information by the scheduler
	PodVGPUPreAllocAnnotation = DomainPrefix + "/pre-allocated"
	// PodVGPURealAllocAnnotation Real device information allocated by device plugins
	PodVGPURealAllocAnnotation = DomainPrefix + "/real-allocated"

	HundredCore     = 100
	MaxDeviceNumber = 16

	// MaxContainerLimit max container num
	MaxContainerLimit = 300000
	// PodAnnotationMaxLength pod annotation max data length 1MB
	PodAnnotationMaxLength = 1024 * 1024

	AllocateCheckErrMsg           = "Allocate check failed"
	PreStartContainerCheckErrMsg  = "PreStartContainer check failed"
	PreStartContainerCheckErrType = "PreStartContainerCheckErr"
)
View Source 
const (
	// CUDA_MEM_LIMIT_<index> gpu memory limit
	CudaMemoryLimitEnv = "CUDA_MEM_LIMIT"
	// CUDA_MEM_RATIO_<index> gpu memory ratio
	CudaMemoryRatioEnv = "CUDA_MEM_RATIO"
	// CUDA_CORE_LIMIT_<index> gpu core limit
	CudaCoreLimitEnv = "CUDA_CORE_LIMIT"
	// CUDA_CORE_SOFT_LIMIT_<index> gpu core soft limit
	CudaSoftCoreLimitEnv = "CUDA_CORE_SOFT_LIMIT"
	// CUDA_CORE_SOFT_LIMIT_<index> gpu memory oversold switch
	CudaMemoryOversoldEnv = "CUDA_MEM_OVERSOLD"
	// GPU_DEVICES_UUID gpu uuid list
	GPUDevicesUuidEnv = "GPU_DEVICES_UUID"
	// CompatibilityModeEnv Indicate the compatibility mode of the environment
	CompatibilityModeEnv = "ENV_COMPATIBILITY_MODE"

	PodNameEnv      = "VGPU_POD_NAME"
	PodNamespaceEnv = "VGPU_POD_NAMESPACE"
	PodUIDEnv       = "VGPU_POD_UID"
	ContNameEnv     = "VGPU_CONTAINER_NAME"
)
View Source 
const (
	DeviceListStrategyEnvvar         = "envvar"
	DeviceListStrategyVolumeMounts   = "volume-mounts"
	DeviceListStrategyCDIAnnotations = "cdi-annotations"
	DeviceListStrategyCDICRI         = "cdi-cri"
)

Constants to represent the various device list strategies

Variables

This section is empty.

Functions

func CheckDeviceType 

func CheckDeviceType(annotations map[string]string, deviceType string) bool

CheckDeviceType Check if the device type meets expectations.

func CheckDeviceUuid 

func CheckDeviceUuid(annotations map[string]string, deviceUUID string) bool

CheckDeviceUuid Check if the device uuid meets expectations.

func FilterAllocatingPods 

func FilterAllocatingPods(activePods []corev1.Pod) []corev1.Pod

FilterAllocatingPods filter out the list of pods to be allocated.

func GetAllocatableOfNode 

func GetAllocatableOfNode(node *corev1.Node, resourceName string) int

GetAllocatableOfNode Return the number of resources that can be allocated to the node.

func GetCapacityOfNode 

func GetCapacityOfNode(node *corev1.Node, resourceName string) int

GetCapacityOfNode Return the capacity of node resources.

func GetContainerRuntime 

func GetContainerRuntime(pod *corev1.Pod, containerName string) (runtimeName string, containerId string)

func GetContainerStatus 

func GetContainerStatus(pod *corev1.Pod, containerName string) (*corev1.ContainerStatus, bool)

func GetCurrentPodByAllocatingPods 

func GetCurrentPodByAllocatingPods(allocatingPods []corev1.Pod) (*corev1.Pod, error)

GetCurrentPodByAllocatingPods find the oldest Pod from the allocating Pods to be allocated as the current Pod to be allocated.

func GetK8sPodCGroupFullPath 

func GetK8sPodCGroupFullPath(podCGroupPath string) string

GetK8sPodCGroupFullPath Obtain the cgroupv2 full path of the pod.

func GetK8sPodCGroupPath 

func GetK8sPodCGroupPath(pod *corev1.Pod) (string, error)

GetK8sPodCGroupPath Obtain the relative path of pod cgroup for k8s.

func GetK8sPodContainerCGroupFullPath 

func GetK8sPodContainerCGroupFullPath(pod *corev1.Pod, containerName string,
	getFullPath func(string) string) (string, error)

func GetK8sPodDeviceCGroupFullPath 

func GetK8sPodDeviceCGroupFullPath(podCGroupPath string) string

GetK8sPodDeviceCGroupFullPath Obtain the full path of the cgroup device subsystem of the pod.

func GetNumaInformation 

func GetNumaInformation(idx int) (int, error)

func GetPredicateTimeOfPod 

func GetPredicateTimeOfPod(pod corev1.Pod) uint64

func GetResourceOfContainer 

func GetResourceOfContainer(container *corev1.Container, resourceName corev1.ResourceName) int

GetResourceOfContainer Return the number of resource limit.

func GetResourceOfPod 

func GetResourceOfPod(pod *corev1.Pod, resourceName corev1.ResourceName) int

GetResourceOfPod Return the number of resource limit for all containers of Pod.

func HasAnnotation 

func HasAnnotation(obj metav1.Object, anno string) (val string, ok bool)

func HasLabel 

func HasLabel(obj metav1.Object, label string) (val string, ok bool)

func InitializeCGroupDriver 

func InitializeCGroupDriver(cgroupDriver string)

func InsertAnnotation 

func InsertAnnotation(obj metav1.Object, k, v string)

func IsShouldDeletePod 

func IsShouldDeletePod(pod *corev1.Pod) bool

IsShouldDeletePod Determine whether the pod has been deleted or needs to be deleted.

func IsVGPUEnabledNode 

func IsVGPUEnabledNode(node *corev1.Node) bool

IsVGPUEnabledNode Determine whether there are VGPU devices on the node.

func IsVGPURequiredContainer 

func IsVGPURequiredContainer(c *corev1.Container) bool

IsVGPURequiredContainer tell if the container is a vGPU request container.

func IsVGPUResourcePod 

func IsVGPUResourcePod(pod *corev1.Pod) bool

IsVGPUResourcePod Determine if a pod has vGPU resource request.

func ParseContainerRuntime 

func ParseContainerRuntime(podContainerId string) (runtimeName string, containerId string)

func PathIsNotExist 

func PathIsNotExist(fullPath string) bool

func PodIsTerminated 

func PodIsTerminated(pod *corev1.Pod) bool

func ShouldRetry 

func ShouldRetry(err error) bool

ShouldRetry Determine whether the error of apiserver is of the type that needs to be retried.

func SplitK8sCGroupBasePath 

func SplitK8sCGroupBasePath(cgroupFullPath string) string

func SystemdPathPrefixOfRuntime 

func SystemdPathPrefixOfRuntime(runtimeName string) string

Types

type AssignedPhase 

type AssignedPhase string
const (
	AssignPhaseSucceed    AssignedPhase = "succeed"
	AssignPhaseAllocating AssignedPhase = "allocating"
	AssignPhaseFailed     AssignedPhase = "failed"
)

type CGroupDriver 

type CGroupDriver string

type CgroupName 

type CgroupName []string

func NewPodCgroupName 

func NewPodCgroupName(pod *corev1.Pod) CgroupName

func (CgroupName) ToCgroupfs 

func (cgroupName CgroupName) ToCgroupfs() string

func (CgroupName) ToSystemd 

func (cgroupName CgroupName) ToSystemd() string

cgroupName.ToSystemd converts the internal cgroup name to a systemd name. For example, the name {"kubepods", "burstable", "pod1234-abcd-5678-efgh"} becomes "/kubepods.slice/kubepods-burstable.slice/kubepods-burstable-pod1234_abcd_5678_efgh.slice" This function always expands the systemd name into the cgroupfs form. If only the last part is needed, use path.Base(...) on it to discard the rest.

type ComputePolicy 

type ComputePolicy string
const (
	// NoneComputePolicy There are no computing power limitations, tasks compete for GPUs on their own.
	NoneComputePolicy ComputePolicy = "none"
	// BalanceComputePolicy Automatically balance GPU load, maximize GPU utilization,
	// allocate idle computing power to ongoing tasks,
	// and roll back excess computing power when new tasks require it.
	BalanceComputePolicy ComputePolicy = "balance"
	// FixedComputePolicy Run tasks with a fixed computing power quota,
	// and the utilization rate will not exceed the quota.
	// Default strategy
	FixedComputePolicy ComputePolicy = "fixed" // default
)

type PodsOrderedByPredicateTime 

type PodsOrderedByPredicateTime []corev1.Pod

func (PodsOrderedByPredicateTime) Len 

func (pods PodsOrderedByPredicateTime) Len() int

func (PodsOrderedByPredicateTime) Less 

func (pods PodsOrderedByPredicateTime) Less(i, j int) bool

func (PodsOrderedByPredicateTime) Swap 

func (pods PodsOrderedByPredicateTime) Swap(i, j int)

type SchedulerPolicy 

type SchedulerPolicy string
const (
	NonePolicy SchedulerPolicy = "none"
	// BinpackPolicy means the lower device memory remained after this allocation, the better
	BinpackPolicy SchedulerPolicy = "binpack"
	// SpreadPolicy means better put this task into an idle GPU card than a shared GPU card
	SpreadPolicy SchedulerPolicy = "spread"
)

type TopologyMode added in v0.3.0 

type TopologyMode string
const (
	// NoneTopology Do not use any topology mode to allocate devices.
	NoneTopology TopologyMode = "none"
	// NUMATopology aligns the allocated devices according to numa nodes.
	NUMATopology TopologyMode = "numa"
	// LinkTopology find the best device set based on link topology.
	LinkTopology TopologyMode = "link"
)

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.