nvidia

README

This package provides an implementation of nvidia device plugin

Behavior

Nvidia device plugin uses NVML bindings to get data regarding available nvidia devices and will expose them via Fingerprint RPC. GPUs can be excluded from fingerprinting by setting the ignored_gpu_ids field. Plugin sends statistics for fingerprinted devices every stats_period period.

Config

The configuration should be passed via an HCL file that begins with a top level config stanza:

config {
  ignored_gpu_ids = ["uuid1", "uuid2"]
  fingerprint_period = "5s"
}

The valid configuration options are:

• ignored_gpu_ids (list(string): []): list of GPU UUIDs strings that should not be exposed to nomad
• fingerprint_period (string: "1m"): interval to repeat the fingerprint process to identify possible changes.

Documentation

Index

• Constants
• Variables
• type Config
• type NvidiaDevice
  • func NewNvidiaDevice(_ context.Context, log log.Logger) *NvidiaDevice
  • func (d *NvidiaDevice) ConfigSchema() (*hclspec.Spec, error)
  • func (d *NvidiaDevice) Fingerprint(ctx context.Context) (<-chan *device.FingerprintResponse, error)
  • func (d *NvidiaDevice) PluginInfo() (*base.PluginInfoResponse, error)
  • func (d *NvidiaDevice) Reserve(deviceIDs []string) (*device.ContainerReservation, error)
  • func (d *NvidiaDevice) SetConfig(cfg *base.Config) error
  • func (d *NvidiaDevice) Stats(ctx context.Context, interval time.Duration) (<-chan *device.StatsResponse, error)

Constants

const (
	// Attribute names and units for reporting Fingerprint output
	MemoryAttr          = "memory"
	PowerAttr           = "power"
	BAR1Attr            = "bar1"
	DriverVersionAttr   = "driver_version"
	CoresClockAttr      = "cores_clock"
	MemoryClockAttr     = "memory_clock"
	PCIBandwidthAttr    = "pci_bandwidth"
	DisplayStateAttr    = "display_state"
	PersistenceModeAttr = "persistence_mode"
)

const (
	// Attribute names for reporting stats output
	PowerUsageAttr = "Power usage"
	PowerUsageUnit = "W"
	PowerUsageDesc = "Power usage for this GPU in watts and " +
		"its associated circuitry (e.g. memory) / Maximum GPU Power"
	GPUUtilizationAttr = "GPU utilization"
	GPUUtilizationUnit = "%"
	GPUUtilizationDesc = "Percent of time over the past sample period " +
		"during which one or more kernels were executing on the GPU."
	MemoryUtilizationAttr  = "Memory utilization"
	MemoryUtilizationUnit  = "%"
	MemoryUtilizationDesc  = "Percentage of bandwidth used during the past sample period"
	EncoderUtilizationAttr = "Encoder utilization"
	EncoderUtilizationUnit = "%"
	EncoderUtilizationDesc = "Percent of time over the past sample period " +
		"during which GPU Encoder was used"
	DecoderUtilizationAttr = "Decoder utilization"
	DecoderUtilizationUnit = "%"
	DecoderUtilizationDesc = "Percent of time over the past sample period " +
		"during which GPU Decoder was used"
	TemperatureAttr      = "Temperature"
	TemperatureUnit      = "C" // Celsius degrees
	TemperatureDesc      = "Temperature of the Unit"
	MemoryStateAttr      = "Memory state"
	MemoryStateUnit      = "MiB" // Mebibytes
	MemoryStateDesc      = "UsedMemory / TotalMemory"
	BAR1StateAttr        = "BAR1 buffer state"
	BAR1StateUnit        = "MiB" // Mebibytes
	BAR1StateDesc        = "UsedBAR1 / TotalBAR1"
	ECCErrorsL1CacheAttr = "ECC L1 errors"
	ECCErrorsL1CacheUnit = "#" // number of errors
	ECCErrorsL1CacheDesc = "Requested L1Cache error counter for the device"
	ECCErrorsL2CacheAttr = "ECC L2 errors"
	ECCErrorsL2CacheUnit = "#" // number of errors
	ECCErrorsL2CacheDesc = "Requested L2Cache error counter for the device"
	ECCErrorsDeviceAttr  = "ECC memory errors"
	ECCErrorsDeviceUnit  = "#" // number of errors
	ECCErrorsDeviceDesc  = "Requested memory error counter for the device"
)

const (
	// Nvidia-container-runtime environment variable names
	NvidiaVisibleDevices = "NVIDIA_VISIBLE_DEVICES"
)

Variables

var (
	// PluginID is the nvidia plugin metadata registered in the plugin
	// catalog.
	PluginID = loader.PluginID{
		Name:       pluginName,
		PluginType: base.PluginTypeDevice,
	}

	// PluginConfig is the nvidia factory function registered in the
	// plugin catalog.
	PluginConfig = &loader.InternalPluginConfig{
		Factory: func(ctx context.Context, l log.Logger) interface{} { return NewNvidiaDevice(ctx, l) },
	}
)

Types

type Config

type Config struct {
	IgnoredGPUIDs     []string `codec:"ignored_gpu_ids"`
	FingerprintPeriod string   `codec:"fingerprint_period"`
}

Config contains configuration information for the plugin.

type NvidiaDevice

type NvidiaDevice struct {
	// contains filtered or unexported fields
}

NvidiaDevice contains all plugin specific data

func NewNvidiaDevice

func NewNvidiaDevice(_ context.Context, log log.Logger) *NvidiaDevice

NewNvidiaDevice returns a new nvidia device plugin.

func (*NvidiaDevice) ConfigSchema

func (d *NvidiaDevice) ConfigSchema() (*hclspec.Spec, error)

ConfigSchema returns the plugins configuration schema.

func (*NvidiaDevice) Fingerprint

func (d *NvidiaDevice) Fingerprint(ctx context.Context) (<-chan *device.FingerprintResponse, error)

Fingerprint streams detected devices. If device changes are detected or the devices health changes, messages will be emitted.

func (*NvidiaDevice) PluginInfo

func (d *NvidiaDevice) PluginInfo() (*base.PluginInfoResponse, error)

PluginInfo returns information describing the plugin.

func (*NvidiaDevice) Reserve

func (d *NvidiaDevice) Reserve(deviceIDs []string) (*device.ContainerReservation, error)

Reserve returns information on how to mount given devices. Assumption is made that nomad server is responsible for correctness of GPU allocations, handling tricky cases such as double-allocation of single GPU

func (*NvidiaDevice) SetConfig

func (d *NvidiaDevice) SetConfig(cfg *base.Config) error

SetConfig is used to set the configuration of the plugin.

func (*NvidiaDevice) Stats

func (d *NvidiaDevice) Stats(ctx context.Context, interval time.Duration) (<-chan *device.StatsResponse, error)

Stats streams statistics for the detected devices.