utils

Documentation

Index

• Constants
• Variables
• func GetAKSGPUImageSHA(size string) string
• func GetGPUDriverVersion(size string) string
• func GetSubnetResourceID(subscriptionID, resourceGroupName, virtualNetworkName, subnetName string) string
• func GetVMName(providerID string) (string, error)
• func GetVnetSubnetIDComponents(vnetSubnetID string) (vnetSubnetResource, error)
• func IsMarinerEnabledGPUSKU(vmSize string) bool
• func IsNvidiaEnabledSKU(vmSize string) bool
• func MkVMID(resourceGroupName string, vmName string) string
• func ResourceIDToProviderID(ctx context.Context, id string) string

Constants

const (
	Nvidia470CudaDriverVersion = "cuda-470.82.01"
	Nvidia550CudaDriverVersion = "cuda-550.54.15"
	Nvidia535GridDriverVersion = "grid-535.161.08"

	// These SHAs will change once we update aks-gpu images in aks-gpu repository. We do that fairly rarely at this time.
	// So for now these will be kept here like this and periodically bump them
	AKSGPUGridSHA = "sha-d1f0ca"
	AKSGPUCudaSHA = "sha-2d4c96"
)

TODO: Get these from agentbaker

Variables

var (
	/* If a new GPU sku becomes available, add a key to this map, but only if you have a confirmation
	   that we have an agreement with NVIDIA for this specific gpu.
	*/
	NvidiaEnabledSKUs = map[string]bool{

		"standard_nv6":      true,
		"standard_nv12":     true,
		"standard_nv12s_v3": true,
		"standard_nv24":     true,
		"standard_nv24s_v3": true,
		"standard_nv24r":    true,
		"standard_nv48s_v3": true,

		"standard_nd6s":   true,
		"standard_nd12s":  true,
		"standard_nd24s":  true,
		"standard_nd24rs": true,

		"standard_nc6s_v2":   true,
		"standard_nc12s_v2":  true,
		"standard_nc24s_v2":  true,
		"standard_nc24rs_v2": true,

		"standard_nc6s_v3":   true,
		"standard_nc12s_v3":  true,
		"standard_nc24s_v3":  true,
		"standard_nc24rs_v3": true,
		"standard_nd40s_v3":  true,
		"standard_nd40rs_v2": true,

		"standard_nc4as_t4_v3":  true,
		"standard_nc8as_t4_v3":  true,
		"standard_nc16as_t4_v3": true,
		"standard_nc64as_t4_v3": true,

		"standard_nd96asr_v4":       true,
		"standard_nd112asr_a100_v4": true,
		"standard_nd120asr_a100_v4": true,

		"standard_nd96amsr_a100_v4":  true,
		"standard_nd112amsr_a100_v4": true,
		"standard_nd120amsr_a100_v4": true,

		"standard_nc24ads_a100_v4": true,
		"standard_nc48ads_a100_v4": true,
		"standard_nc96ads_a100_v4": true,
		"standard_ncads_a100_v4":   true,

		"standard_nc8ads_a10_v4":  true,
		"standard_nc16ads_a10_v4": true,
		"standard_nc32ads_a10_v4": true,

		"standard_nv6ads_a10_v5":   true,
		"standard_nv12ads_a10_v5":  true,
		"standard_nv18ads_a10_v5":  true,
		"standard_nv36ads_a10_v5":  true,
		"standard_nv36adms_a10_v5": true,
		"standard_nv72ads_a10_v5":  true,

		"standard_nd96ams_v4":      true,
		"standard_nd96ams_a100_v4": true,
	}

	// List of GPU SKUs currently enabled and validated for Mariner. Will expand the support
	// to cover other SKUs available in Azure
	MarinerNvidiaEnabledSKUs = map[string]bool{

		"standard_nc6s_v3":   true,
		"standard_nc12s_v3":  true,
		"standard_nc24s_v3":  true,
		"standard_nc24rs_v3": true,
		"standard_nd40s_v3":  true,
		"standard_nd40rs_v2": true,

		"standard_nc4as_t4_v3":  true,
		"standard_nc8as_t4_v3":  true,
		"standard_nc16as_t4_v3": true,
		"standard_nc64as_t4_v3": true,
	}
)

var ConvergedGPUDriverSizes = map[string]bool{
	"standard_nv6ads_a10_v5":   true,
	"standard_nv12ads_a10_v5":  true,
	"standard_nv18ads_a10_v5":  true,
	"standard_nv36ads_a10_v5":  true,
	"standard_nv72ads_a10_v5":  true,
	"standard_nv36adms_a10_v5": true,
	"standard_nc8ads_a10_v4":   true,
	"standard_nc16ads_a10_v4":  true,
	"standard_nc32ads_a10_v4":  true,
}

ConvergedGPUDriverSizes : these sizes use a "converged" driver to support both cuda/grid workloads.

how do you figure this out? ask HPC or find out by trial and error. installing vanilla cuda drivers will fail to install with opaque errors. see https://github.com/Azure/azhpc-extensions/blob/daaefd78df6f27012caf30f3b54c3bd6dc437652/NvidiaGPU/resources.json

Functions

func GetAKSGPUImageSHA

func GetAKSGPUImageSHA(size string) string

func GetGPUDriverVersion

func GetGPUDriverVersion(size string) string

NV series GPUs target graphics workloads vs NC which targets compute. they typically use GRID, not CUDA drivers, and will fail to install CUDA drivers. NVv1 seems to run with CUDA, NVv5 requires GRID. NVv3 is untested on AKS, NVv4 is AMD so n/a, and NVv2 no longer seems to exist (?).

func GetSubnetResourceID

func GetSubnetResourceID(subscriptionID, resourceGroupName, virtualNetworkName, subnetName string) string

GetSubnetResourceID constructs the subnet resource id

func GetVMName

func GetVMName(providerID string) (string, error)

GetVMName parses the provider ID stored on the node to get the vmName associated with a node

func GetVnetSubnetIDComponents

func GetVnetSubnetIDComponents(vnetSubnetID string) (vnetSubnetResource, error)

func IsMarinerEnabledGPUSKU

func IsMarinerEnabledGPUSKU(vmSize string) bool

IsNvidiaEnabledSKU determines if an VM SKU has nvidia driver support

func IsNvidiaEnabledSKU

func IsNvidiaEnabledSKU(vmSize string) bool

IsNvidiaEnabledSKU determines if an VM SKU has nvidia driver support

func MkVMID

func MkVMID(resourceGroupName string, vmName string) string

func ResourceIDToProviderID

func ResourceIDToProviderID(ctx context.Context, id string) string