gpu

Documentation

Index

• Constants
• func FilterOutNodesWithUnreadyGpus(allNodes, readyNodes []*apiv1.Node) ([]*apiv1.Node, []*apiv1.Node)
• func GetGpuTypeForMetrics(node *apiv1.Node, nodeGroup cloudprovider.NodeGroup) string
• func GetNodeTargetGpus(node *apiv1.Node, nodeGroup cloudprovider.NodeGroup) (gpuType string, gpuCount int64, error errors.AutoscalerError)
• func NodeHasGpu(node *apiv1.Node) bool
• func PodRequestsGpu(pod *apiv1.Pod) bool

Constants

const (
	// ResourceNvidiaGPU is the name of the Nvidia GPU resource.
	ResourceNvidiaGPU = "nvidia.com/gpu"
	// GPULabel is the label added to nodes with GPU resource on GKE.
	GPULabel = "cloud.google.com/gke-accelerator"
	// DefaultGPUType is the type of GPU used in NAP if the user
	// don't specify what type of GPU his pod wants.
	DefaultGPUType = "nvidia-tesla-k80"
)

const (
	// MetricsGenericGPU - for when there is no information about GPU type
	MetricsGenericGPU = "generic"
	// MetricsMissingGPU - for when there's a label, but GPU didn't appear
	MetricsMissingGPU = "missing-gpu"
	// MetricsUnexpectedLabelGPU - for when there's a label, but no GPU at all
	MetricsUnexpectedLabelGPU = "unexpected-label"
	// MetricsUnknownGPU - for when GPU type is unknown
	MetricsUnknownGPU = "not-listed"
	// MetricsErrorGPU - for when there was an error obtaining GPU type
	MetricsErrorGPU = "error"
	// MetricsNoGPU - for when there is no GPU and no label all
	MetricsNoGPU = ""
)

Functions

func FilterOutNodesWithUnreadyGpus

func FilterOutNodesWithUnreadyGpus(allNodes, readyNodes []*apiv1.Node) ([]*apiv1.Node, []*apiv1.Node)

FilterOutNodesWithUnreadyGpus removes nodes that should have GPU, but don't have it in allocatable from ready nodes list and updates their status to unready on all nodes list. This is a hack/workaround for nodes with GPU coming up without installed drivers, resulting in GPU missing from their allocatable and capacity.

func GetGpuTypeForMetrics

func GetGpuTypeForMetrics(node *apiv1.Node, nodeGroup cloudprovider.NodeGroup) string

GetGpuTypeForMetrics returns name of the GPU used on the node or empty string if there's no GPU if the GPU type is unknown, "generic" is returned NOTE: current implementation is GKE/GCE-specific

func GetNodeTargetGpus

func GetNodeTargetGpus(node *apiv1.Node, nodeGroup cloudprovider.NodeGroup) (gpuType string, gpuCount int64, error errors.AutoscalerError)

GetNodeTargetGpus returns the number of gpus on a given node. This includes gpus which are not yet ready to use and visible in kubernetes.

func NodeHasGpu

func NodeHasGpu(node *apiv1.Node) bool

NodeHasGpu returns true if a given node has GPU hardware. The result will be true if there is hardware capability. It doesn't matter if the drivers are installed and GPU is ready to use.

func PodRequestsGpu

func PodRequestsGpu(pod *apiv1.Pod) bool

PodRequestsGpu returns true if a given pod has GPU request.