What Can I Do If a GPU Card Is Unavailable on a GPU Node?

1. Check the nvidia-driver-installer pod status.
   Log in to the CCE console and click the cluster name to access the cluster Overview page. In the navigation pane, choose Nodes. In the right pane, click the Nodes tab. Locate the row containing the target node, choose More > Pods in the Operation column, and check whether the nvidia-driver-installer pod runs on the node. If the nvidia-driver-installer pod is present and is:

   • In the Running state: The pod is functioning properly. Proceed to 2 to verify whether the driver was installed.
   • Not in the Running state for an extended period: Check the pod events for any abnormalities and troubleshoot based on the reported error information. If the fault persists, submit a service ticket and contact technical support.

   The name of the nvidia-driver-installer pod varies depending on the OS. The details are listed in the table below.

   Table 1 Names of the nvidia-driver-installer pod

   OS	Pod Name
   Huawei Cloud EulerOS 2.0	hce20-nvidia-driver-installer
   Ubuntu	ubuntu22-nvidia-driver-installer
   Others	nvidia-driver-installer

2. Check whether the GPU driver has been installed.
   1. In the node list, click the name of the target node. In the dialog box displayed, click OK. On the node details page, click Remote Login in the upper right corner.
   2. Check the driver installation directory.
      1. Check whether the directory exists. If it is present, run the below command to go to the driver installation directory. If it is not present, skip this step and go to 3 to check whether there is an error during the driver installation.

         cd <Driver installation directory>

         The driver installation directory varies depending on the CCE AI Suite (NVIDIA GPU) add-on version. The details are as follows:

         • If the CCE AI Suite (NVIDIA GPU) add-on version is later than 2.0.0, the driver installation directory is /usr/local/nvidia.
         • If the CCE AI Suite (NVIDIA GPU) add-on version is earlier than 2.0.0, the driver installation directory is /opt/cloud/cce/nvidia.
      2. Run the following command in the driver installation directory to view all files in the directory:

         ls -l

         The figure below shows a typical file directory. nvidia.run is the driver installation file. nvidia-installer.log is the installation logs generated by the NVIDIA driver. nvidia-uninstall.log, if present, is the corresponding uninstallation logs, though it may not always appear in the directory. If any files are missing, except for nvidia-uninstall.log, go to 3 to check whether there is an error during the driver installation.

         

      3. Run the below command to go to the bin directory of NVIDIA and check whether nvidia-smi is functioning properly. If the add-on version is earlier than 2.0.0, replace the path with opt/cloud/cce/nvidia/bin.

         cd /usr/local/nvidia/bin
         ./nvidia-smi

         If information similar to that shown in the figure below is not displayed, go to 3 to check whether there is an error during the driver installation.

         

3. View the node driver installation logs to check whether there is an error during the driver installation.
   Run the below command to view the logs of the nvidia-driver-installer pod. If the add-on version is earlier than 2.0.0, replace the path with /opt/cloud/cce/nvidia/nvidia-installer.log.

   cat /usr/local/nvidia/nvidia-installer.log

   If the command output contains the below information, the driver installation completed without error. Otherwise, an error occurred during the installation. If there is an error, submit a service ticket and contact technical support.

   ...
   > Installation of the NVIDIA Accelerated Graphics Driver for xxx (version: x.x.x) is now complete.

In a CCE cluster, device-plugin is responsible for reporting hardware resource statuses. In GPU scenarios, nvidia-gpu-device-plugin in the kube-system namespace reports the available GPU resources on each node. If the reported GPU resources appear incorrect or if device mounting issues occur, it is advised to first check device-plugin for potential anomalies.

kubectl get po -A -owide|grep nvidia

• If the device-plugin pod is not in the Running state, submit a service ticket and contact technical support.
• If the device-plugin pod is in the Running state, run the following command to check its logs for errors:

  kubectl logs -n kube-system nvidia-gpu-device-plugin-9xmhr

  If "gpu driver wasn't ready. will re-check" is displayed in the command output, go to 2 and check whether the /usr/local/nvidia/bin/nvidia-smi or /opt/cloud/cce/nvidia/bin/nvidia-smi file exists in the driver installation directory.

  ...
  I0527 11:29:06.420714 3336959 nvidia_gpu.go:76] device-plugin started
  I0527 11:29:06.521884 3336959 nodeinformer.go:124] "nodeInformer started"
  I0527 11:29:06.521964 3336959 nvidia_gpu.go:262] "gpu driver wasn't ready. will re-check in %s" 5s="(MISSING)"
  I0527 11:29:11.524882 3336959 nvidia_gpu.go:262] "gpu driver wasn't ready. will re-check in %s" 5s="(MISSING)"
  ...

Rectify the fault by referring to GPU Fault Handling.