Nvidia GPU Device Plugin

Name: nomad-device-nvidia

Use the NVIDIA device plugin to expose NVIDIA GPUs to Nomad. The driver automatically supports Multi-Instance GPU (MIG).

The NVIDIA device plugin uses NVML bindings to get data regarding available NVIDIA devices and then exposes them via Fingerprint RPC. The plugin detects whether the GPU has Multi-Instance GPU enabled, and when enabled, the plugin fingerprints all instances as individual GPUs. You may exclude GPUs from fingerprinting by setting the ignored_gpu_ids field.

Fingerprinted Attributes

Attribute	Unit
memory	MiB
power	W (Watt)
bar1	MiB
driver_version	string
cores_clock	MHz
memory_clock	MHz
pci_bandwidth	MB/s
display_state	string
persistence_mode	string

Runtime Environment

The nvidia-gpu device plugin exposes the following environment variables:

NVIDIA_VISIBLE_DEVICES - List of NVIDIA GPU IDs available to the task.

Additional Task Configurations

Additional environment variables can be set by the task to influence the runtime environment. See Nvidia's documentation.

Installation Requirements

In order to use the nomad-device-nvidia device driver the following prerequisites must be met:

GNU/Linux x86_64 with kernel version > 3.10
NVIDIA GPU with Architecture > Fermi (2.1)
NVIDIA drivers >= 340.29 with binary nvidia-smi
Docker v19.03+

Container Toolkit Installation

Follow the NVIDIA Container Toolkit installation instructions from NVIDIA to prepare a machine to use docker containers with NVIDIA GPUs. You should be able to run this simple command to test your environment and produce meaningful output.

docker run --rm --gpus all nvidia/cuda:11.0-base nvidia-smi

Plugin Configuration

plugin "nomad-device-nvidia" {  config {    enabled            = true    ignored_gpu_ids    = ["GPU-fef8089b", "GPU-ac81e44d"]    fingerprint_period = "1m"  }}

The nomad-device-nvidia device plugin supports the following configuration in the agent config:

enabled (bool: true) - Control whether the plugin should be enabled and running.
ignored_gpu_ids (array<string>: []) - Specifies the set of GPU UUIDs that should be ignored when fingerprinting.
fingerprint_period (string: "1m") - The period in which to fingerprint for device changes.

Limitations

The NVIDIA integration only works with drivers who natively integrate with NVIDIA's container runtime library.

Nomad has tested support with the docker driver.

Source Code & Compiled Binaries

The source code for this plugin can be found at hashicorp/nomad-device-nvidia. You can also find pre-built binaries on the releases page.

Examples

Inspect a node with a GPU:

$ nomad node status 4d46e59f // ...TRUNCATED... Device Resource Utilizationnvidia/gpu/Tesla K80[GPU-e1f6f4f1-1ea5-7b9d-5f03-338a9dc32416]  0 / 11441 MiB

Display detailed statistics on a node with a GPU:

$ nomad node status -stats 4d46e59f // ...TRUNCATED... Device Resource Utilizationnvidia/gpu/Tesla K80[GPU-e1f6f4f1-1ea5-7b9d-5f03-338a9dc32416]  0 / 11441 MiB // ...TRUNCATED... Device StatsDevice              = nvidia/gpu/Tesla K80[GPU-e1f6f4f1-1ea5-7b9d-5f03-338a9dc32416]BAR1 buffer state   = 2 / 16384 MiBDecoder utilization = 0 %ECC L1 errors       = 0ECC L2 errors       = 0ECC memory errors   = 0Encoder utilization = 0 %GPU utilization     = 0 %Memory state        = 0 / 11441 MiBMemory utilization  = 0 %Power usage         = 37 / 149 WTemperature         = 34 C

Run the following example job to see that the GPU was mounted in the container:

job "gpu-test" {  datacenters = ["dc1"]  type = "batch"   group "smi" {    task "smi" {      driver = "docker"       config {        image = "nvidia/cuda:11.0-base"        command = "nvidia-smi"      }       resources {        device "nvidia/gpu" {          count = 1           # Add an affinity for a particular model          affinity {            attribute = "${device.model}"            value     = "Tesla K80"            weight    = 50          }        }      }    }  }}

$ nomad run example.nomad.hcl==> Monitoring evaluation "21bd7584"    Evaluation triggered by job "gpu-test"    Allocation "d250baed" created: node "4d46e59f", group "smi"    Evaluation status changed: "pending" -> "complete"==> Evaluation "21bd7584" finished with status "complete" $ nomad alloc status d250baed // ...TRUNCATED... Task "smi" is "dead"Task ResourcesCPU        Memory       Disk     Addresses0/100 MHz  0 B/300 MiB  300 MiB Device Statsnvidia/gpu/Tesla K80[GPU-e1f6f4f1-1ea5-7b9d-5f03-338a9dc32416]  0 / 11441 MiB Task Events:Started At     = 2019-01-23T18:25:32ZFinished At    = 2019-01-23T18:25:34ZTotal Restarts = 0Last Restart   = N/A Recent Events:Time                  Type        Description2019-01-23T18:25:34Z  Terminated  Exit Code: 02019-01-23T18:25:32Z  Started     Task started by client2019-01-23T18:25:29Z  Task Setup  Building Task Directory2019-01-23T18:25:29Z  Received    Task received by client $ nomad alloc logs d250baedWed Jan 23 18:25:32 2019+-----------------------------------------------------------------------------+| NVIDIA-SMI 410.48                 Driver Version: 410.48                    ||-------------------------------+----------------------+----------------------+| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC || Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. ||===============================+======================+======================||   0  Tesla K80           On   | 00004477:00:00.0 Off |                    0 || N/A   33C    P8    37W / 149W |      0MiB / 11441MiB |      0%      Default |+-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+| Processes:                                                       GPU Memory ||  GPU       PID   Type   Process name                             Usage      ||=============================================================================||  No running processes found                                                 |+-----------------------------------------------------------------------------+