nvidia

Documentation

Index

• Constants
• Variables
• func NewNvidiaTree(cfg *config.Config) device.GPUTree
• type DeviceMeta
• type LessFunc
• type LevelMap
• type NvidiaNode
  • func NewNvidiaNode(t *NvidiaTree) *NvidiaNode
  • func (n *NvidiaNode) Available() int
  • func (n *NvidiaNode) GetAvailableLeaves() []*NvidiaNode
  • func (n *NvidiaNode) MinorName() string
  • func (n *NvidiaNode) String() string
  • func (n *NvidiaNode) Type() int
• type NvidiaTree
  • func (t *NvidiaTree) Available() int
  • func (t *NvidiaTree) Init(input string)
  • func (t *NvidiaTree) Leaves() []*NvidiaNode
  • func (t *NvidiaTree) MarkFree(node *NvidiaNode, util int64, memory int64)
  • func (t *NvidiaTree) MarkOccupied(node *NvidiaNode, util int64, memory int64)
  • func (t *NvidiaTree) PrintGraph() string
  • func (t *NvidiaTree) Query(name string) *NvidiaNode
  • func (t *NvidiaTree) Root() *NvidiaNode
  • func (t *NvidiaTree) Total() int
  • func (t *NvidiaTree) Update()
• type SchedulerCache

Constants

const (
	//MaxProcess is the Maximum number of process in one device.
	MaxProcess = 64
	//NamePattern is the name pattern of nvidia device.
	NamePattern = "/dev/nvidia%d"

	//HundredCore represents 100 virtual cores.
	HundredCore = 100
)

Variables

var (
	//ByType compares two NvidiaNode by GpuTopologyLevel
	ByType = func(p1, p2 *NvidiaNode) bool {
		return p1.Type() < p2.Type()
	}

	//ByAvailable compares two NvidiaNode by count of available leaves
	ByAvailable = func(p1, p2 *NvidiaNode) bool {
		return p1.Available() < p2.Available()
	}

	//ByID compares two NvidiaNode by ID
	ByID = func(p1, p2 *NvidiaNode) bool {
		return p1.Meta.ID < p2.Meta.ID
	}

	//ByMemory compares two NvidiaNode by memory already used
	ByMemory = func(p1, p2 *NvidiaNode) bool {
		return p1.Meta.UsedMemory < p2.Meta.UsedMemory
	}

	//ByPids compares two NvidiaNode by length of PIDs running on node
	ByPids = func(p1, p2 *NvidiaNode) bool {
		return len(p1.Meta.Pids) < len(p2.Meta.Pids)
	}

	//ByAllocatableCores compares two NvidiaNode by available cores
	ByAllocatableCores = func(p1, p2 *NvidiaNode) bool {
		return p1.AllocatableMeta.Cores < p2.AllocatableMeta.Cores
	}

	//ByAllocatableMemory compares two NvidiaNode by available memory
	ByAllocatableMemory = func(p1, p2 *NvidiaNode) bool {
		return p1.AllocatableMeta.Memory/types.MemoryBlockSize < p2.AllocatableMeta.Memory/types.MemoryBlockSize
	}

	//PrintSorter is used to sort nodes when printing them out
	PrintSorter = &printSort{
		less: []LessFunc{ByType, ByAvailable, ByID},
	}
)

Functions

func NewNvidiaTree

func NewNvidiaTree(cfg *config.Config) device.GPUTree

NewNvidiaTree returns a new NvidiaTree.

Types

type DeviceMeta

type DeviceMeta struct {
	ID          int
	MinorID     int
	UsedMemory  uint64
	TotalMemory uint64
	Pids        []uint
	BusId       string
	Utilization uint
	UUID        string
}

DeviceMeta contains metadata of GPU device

type LessFunc

type LessFunc func(p1, p2 *NvidiaNode) bool

LessFunc represents funcion to compare two NvidiaNode

type LevelMap

type LevelMap map[nvml.GpuTopologyLevel][]*NvidiaNode

LevelMap is a map stores NvidiaNode on each level.

type NvidiaNode

type NvidiaNode struct {
	Meta            DeviceMeta
	AllocatableMeta SchedulerCache

	Parent   *NvidiaNode
	Children []*NvidiaNode
	Mask     uint32
	// contains filtered or unexported fields
}

NvidiaNode represents a node of Nvidia GPU

func NewNvidiaNode

func NewNvidiaNode(t *NvidiaTree) *NvidiaNode

NewNvidiaNode returns a new NvidiaNode

func (*NvidiaNode) Available

func (n *NvidiaNode) Available() int

Available returns conut of available leaves of this NvidiaNode.

func (*NvidiaNode) GetAvailableLeaves

func (n *NvidiaNode) GetAvailableLeaves() []*NvidiaNode

GetAvailableLeaves returns leaves of this NvidiaNode which available for allocating.

func (*NvidiaNode) MinorName

func (n *NvidiaNode) MinorName() string

MinorName returns MinorID of this NvidiaNode

func (*NvidiaNode) String

func (n *NvidiaNode) String() string

func (*NvidiaNode) Type

func (n *NvidiaNode) Type() int

Type returns GpuTopologyLevel of this NvidiaNode

type NvidiaTree

type NvidiaTree struct {
	sync.Mutex
	// contains filtered or unexported fields
}

NvidiaTree represents a Nvidia GPU in a tree.

func (*NvidiaTree) Available

func (t *NvidiaTree) Available() int

Available returns number of available leaves of this tree.

func (*NvidiaTree) Init

func (t *NvidiaTree) Init(input string)

Init a NvidiaTree. Will try to use nvml first, fallback to input string if parseFromLibrary() failed.

func (*NvidiaTree) Leaves

func (t *NvidiaTree) Leaves() []*NvidiaNode

Leaves returns leaves of tree

func (*NvidiaTree) MarkFree

func (t *NvidiaTree) MarkFree(node *NvidiaNode, util int64, memory int64)

MarkFree updates a NvidiaNode by freeing request cores and memory. If request cores < HundredCore, plus available cores and memory with request value. If request cores >= HundredCore, set available cores and memory to total, and update mask of all parents of this node.

func (*NvidiaTree) MarkOccupied

func (t *NvidiaTree) MarkOccupied(node *NvidiaNode, util int64, memory int64)

MarkOccupied updates a NvidiaNode by adding request cores and memory. Mask of all parents of this node will be updated. If request cores < HundredCore, minus available cores and memory with request value. If request cores >= HundredCore, set available cores and memory to 0,

func (*NvidiaTree) PrintGraph

func (t *NvidiaTree) PrintGraph() string

PrintGraph returns the details of tree as string

func (*NvidiaTree) Query

func (t *NvidiaTree) Query(name string) *NvidiaNode

Query tries to find node by name, return nil if not found

func (*NvidiaTree) Root

func (t *NvidiaTree) Root() *NvidiaNode

Root returns root node of tree

func (*NvidiaTree) Total

func (t *NvidiaTree) Total() int

Total returns count of leaves

func (*NvidiaTree) Update

func (t *NvidiaTree) Update()

Update NvidiaTree by info getting from GPU devices. Return immediately if real GPU device is not available.

type SchedulerCache

type SchedulerCache struct {
	Cores  int64
	Memory int64
}

SchedulerCache contains allocatable resource of GPU