Node Auto-Provisioning failing for certain GPU nodes (T4)

[agam]

How to re-create

A job that is marked as requiring nvidia.com/gpu, if results in a new node being spun up in GKE, will fail to be scheduled on that node.

Why is this bad

• Using GPU nodes with Node-Auto-Provisioning in GKE is broken (at least for T4s, not sure which other GPU types are affected)
• It feels strange that such a core "elasticity behavior" is unacknowledged -- hoping this issue gets attention and results in at least an ETA for the fix

Details on error

The provisioned node has a nvidia-device-plugin pod
This pod has a nvidia-driver-installer container which is an init container
This container is stuck on startup

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
   0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0 100   720  100   720    0     0   113k      0 --:--:-- --:--:-- --:--:--  117k
GPU driver auto installation is disabled.
Waiting for GPU driver libraries to be available.

As a result, the kubelet never registers the nvidia.com/gpu resource, which means that the job (which triggered the node in the first place!) can't get its pods scheduled on it.

Prior context:

This is based off the following issue, which is no longer fixed (but which I cannot reopen)

#356

[agam]

fwiw I'm seeing this on A100 nodes too now

[jian-mo]

I'm seeing this on T4 as well

[itssubhodiproy]

for l4 gpus, it is also not working. I'm using COS image and didn't disabled automatic installation

failed to get driver version with error: Failed to download and read GPU driver versions proto with error: failed to download gpu_driver_versions.bin from GCS bucket with error: failed to download gpu_driver_versions.bin artifact from bucket: cos-tools, object: 18244.151.14/lakitu/gpu_driver_versions.bin to /root/home/kubernetes/bin/nvidia/gpu_driver_versions.bin with error: failed to create the reader from GCS client: googleapi: got HTTP response code 403 with body: <?xml version='1.0' encoding='UTF-8'?><Error><Code>AccessDenied</Code><Message>Access denied.</Message></Error>

[AssenDimitrov]

Seems like this is still a problem.... T4 in London:

% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 759 100 759 0 0 328k 0 --:--:-- --:--:-- --:--:-- 370k
GPU driver auto installation is disabled.
Waiting for GPU driver libraries to be available

stuck like this forever....

[MeCode4Food]

I'm seeing this for T4 GPU nodes on my end as well in asia-southeast.

GPU driver auto installation is disabled.
Waiting for GPU driver libraries to be available

[MeCode4Food]

> curl -H "Metadata-Flavor:Google" http://metadata.google.internal/computeMetadata/v1/instance/attributes/kube-labels
# prettified for clarity
cloud.google.com/gke-accelerator=nvidia-tesla-t4,
cloud.google.com/gke-boot-disk=pd-balanced,
cloud.google.com/gke-container-runtime=containerd,
cloud.google.com/gke-cpu-scaling-level=2,
cloud.google.com/gke-logging-variant=DEFAULT,
cloud.google.com/gke-max-pods-per-node=110,
cloud.google.com/gke-memory-gb-scaling-level=7,
cloud.google.com/gke-nodepool=nap-n1-standard-2-gpu1-xxxxxxxx,
cloud.google.com/gke-os-distribution=cos,
cloud.google.com/gke-provisioning=standard,
cloud.google.com/gke-stack-type=IPV4,
cloud.google.com/machine-family=n1,
cloud.google.com/private-node=false

From the following init container's command:

...
  initContainers:
  - command:
    - bash
    - -c
    - |
      LABELS=$( curl --retry 5 -H "Metadata-Flavor:Google" http://metadata.google.internal/computeMetadata/v1/instance/attributes/kube-labels || exit 1 )
      IFS=,; for label in $LABELS; do
        IFS==; read -r LABEL VALUE <<< "$label"
        if [[ "${LABEL}" == "cloud.google.com/gke-gpu-driver-version" ]]; then
          GPU_DRIVER_VERSION=$VALUE
        fi
      done
      if [[ "${GPU_DRIVER_VERSION}" == "latest" ]]; then
        echo "latest" > /etc/nvidia/gpu_driver_version_config.txt
        /cos-gpu-installer install --version=latest || exit 1
      elif [[ "${GPU_DRIVER_VERSION}" == "default" ]]; then
        echo "default" > /etc/nvidia/gpu_driver_version_config.txt
        /cos-gpu-installer install || exit 1
      else
        echo "disabled" > /etc/nvidia/gpu_driver_version_config.txt
        echo "GPU driver auto installation is disabled."
      fi
      echo "Waiting for GPU driver libraries to be available."
      while ! [[ -f /usr/local/nvidia/lib64/libcuda.so ]]; do
        sleep 5
      done
      echo "GPU driver is installed."
      echo "InitContainer succeeded. Start nvidia-gpu-device-plugin container."
      exit 0

Seems like the key cloud.google.com/gke-gpu-driver-version is missing from the metadata response

[AssenDimitrov]

Workaround for me was to use Ubuntu OS and manually install the driver with applying their example: https://cloud.google.com/kubernetes-engine/docs/how-to/gpus#installing_drivers

[raushan2016]

Based on the doc here, while enabling Node-Auto-Provisioning the default GPU driver installation mode is disabled.
When creating the workload(pod/deployment) you can add the driver-version in the nodeSelector and the new nodel pool will be created with the right driver version.

spec:
  nodeSelector:
    cloud.google.com/gke-gpu-driver-version: "latest"

https://cloud.google.com/kubernetes-engine/docs/how-to/node-auto-provisioning#choose-driver-version