nvidia

Documentation

Index

• func NewNvidiaGPUManager(devDirectory, procDirectory string, mountPaths []pluginapi.Mount, ...) *nvidiaGPUManager
• func RegisterWithV1Beta1Kubelet(kubeletEndpoint, pluginEndpoint, resourceName string) error
• type GPUConfig
  • func (config *GPUConfig) AddDefaultsAndValidate() error
• type GPUSharingConfig

Functions

func NewNvidiaGPUManager

func NewNvidiaGPUManager(devDirectory, procDirectory string, mountPaths []pluginapi.Mount, gpuConfig GPUConfig) *nvidiaGPUManager

func RegisterWithV1Beta1Kubelet

func RegisterWithV1Beta1Kubelet(kubeletEndpoint, pluginEndpoint, resourceName string) error

TODO: remove this function once we move to probe based registration.

Types

type GPUConfig

type GPUConfig struct {
	GPUPartitionSize string
	// MaxTimeSharedClientsPerGPU is the number of the time-shared GPU resources to expose for each physical GPU.
	// Deprecated in favor of GPUSharingConfig.
	MaxTimeSharedClientsPerGPU int
	// GPUSharingConfig informs how GPUs on this node can be shared between containers.
	GPUSharingConfig GPUSharingConfig
	// Xid error codes that will set the node to unhealthy
	HealthCriticalXid []int
}

GPUConfig stores the settings used to configure the GPUs on a node.

func (*GPUConfig) AddDefaultsAndValidate

func (config *GPUConfig) AddDefaultsAndValidate() error

type GPUSharingConfig

type GPUSharingConfig struct {
	// GPUSharingStrategy is the type of sharing strategy to enable on this node. Values are "time-sharing" or "mps".
	GPUSharingStrategy gpusharing.GPUSharingStrategy
	// MaxSharedClientsPerGPU is the maximum number of clients that are allowed to share a single GPU.
	MaxSharedClientsPerGPU int
}