CCE AI Suite (NVIDIA GPU)

Introduction

NVIDIA GPU is a device management add-on that supports GPUs in containers. To use GPU nodes in a cluster, this add-on must be installed.

Constraints

The driver to be downloaded must be a .run file.
Only NVIDIA Tesla drivers are supported, not GRID drivers.
When installing or reinstalling the add-on, ensure that the driver download address is correct and accessible. CCE does not verify the address validity.
The gpu-beta add-on only enables you to download the driver and execute the installation script. The add-on status only indicates that how the add-on is running, not whether the driver is successfully installed.
CCE does not guarantee the compatibility between the GPU driver version and the CDUA library version of your application. You need to check the compatibility by yourself.
If a custom OS image has had a a GPU driver installed, CCE cannot ensure that the GPU driver is compatible with other GPU components such as the monitoring components used in CCE.

Installing the Add-on

Log in to the CCE console and click the cluster name to access the cluster console. Choose Add-ons in the navigation pane, locate CCE AI Suite (NVIDIA GPU) on the right, and click Install.
Configure the add-on parameters.
- NVIDIA Driver: Enter the link for downloading the NVIDIA driver. All GPU nodes in the cluster will use this driver.
  - If the download link is a public network address, for example, https://us.download.nvidia.com/tesla/470.103.01/NVIDIA-Linux-x86_64-470.103.01.run, bind an EIP to each GPU node. For details about how to obtain the driver link, see Obtaining the Driver Link from Public Network.
  - If the download link is an OBS URL, you do not need to bind an EIP to GPU nodes. For details about how to obtain the driver link, see Obtaining the Driver Link from OBS.
  - Ensure that the NVIDIA driver version matches the GPU node.
  - After the driver version is changed, restart the node for the change to take effect.
- Driver Selection: If you do not want all GPU nodes in a cluster to use the same driver, CCE allows you to install a different GPU driver for each node pool.
  - The add-on installs the driver with the version specified by the node pool. The driver takes effect only for new pool nodes.
  - After the driver version is updated, it takes effect on the nodes newly added to the node pool. Existing nodes must restart to apply the changes.
- GPU virtualization (supported in 2.0.5 and later versions): Enable GPU virtualization to support the segmentation and isolation for the compute power and GPU memory of a single GPU.
  If the Volcano add-on has not been installed in the cluster, GPU virtualization cannot be enabled. Click One-click installation to install it. To configure the Volcano add-on parameters during installation, click Custom Installation. For details, see Volcano Scheduler.
  
  If the Volcano add-on has been installed in the cluster but its version does not support GPU virtualization, click Upgrade to upgrade it. To configure the Volcano add-on parameters during installation, click Custom Upgrade. For details, see Volcano Scheduler.
  After GPU virtualization is enabled, select Virtualization nodes are compatible with GPU sharing mode, that is, default GPU scheduling in Kubernetes is supported. This capability requires that the version of gpu-device-plugin is 2.0.10 or later and the version of Volcano is 1.10.5 or later.
  - If you enable compatibility, the nvidia.com/gpu quota specified in workloads (the nvidia.com/gpu quota is set to a decimal fraction, for example, 0.5) is provided by GPU virtualization to implement GPU memory isolation. The GPU memory is allocated to containers based on the specified quota. For example, 8 GiB (0.5 x 16 GiB) GPU memory is allocated. The value of GPU memory must be an integer multiple of 128 MiB. Otherwise, the value is automatically rounded down to the nearest integer. If nvidia.com/gpu resources have been used in the workload before compatibility is enabled, the resources will not be provided by GPU virtualization but the entire GPU.
  - After compatibility is enabled, if you use the nvidia.com/gpu quota, it is equivalent to enabling GPU memory isolation. The nvidia.com/gpu quota can share a GPU with workloads in GPU memory isolation mode, but cannot share a GPU with workloads in compute and GPU memory isolation mode.
  - If compatibility is disabled, the nvidia.com/gpu quota specified in the workload only affects the scheduling result. It does not require GPU memory isolation. That is, although the nvidia.com/gpu quota is set to 0.5, you can still view complete GPU memory in the container. In addition, workloads using nvidia.com/gpu resources and workloads using virtualized GPU memory cannot be scheduled to the same node.
  - If you deselect Virtualization nodes are compatible with GPU sharing mode, running workloads will not be affected, but workloads may fail to be scheduled. For example, if compatibility is disabled, the workload using nvidia.com/gpu resources are still in the GPU memory isolation mode. As a result, the GPU cannot schedule workloads in compute and GPU memory isolation mode. You need to delete workloads using nvidia.com/gpu resources before rescheduling.
Click Install.

If the add-on is uninstalled, GPU pods newly scheduled to the nodes cannot run properly, but GPU pods already running on the nodes will not be affected.

Verifying the Add-on

After the add-on is installed, run the nvidia-smi command on the GPU node and the container that schedules GPU resources to verify the availability of the GPU device and driver.

GPU node:

# If the add-on version is earlier than 2.0.0, run the following command:
cd /opt/cloud/cce/nvidia/bin && ./nvidia-smi

# If the add-on version is 2.0.0 or later and the driver installation path is changed, run the following command:
cd /usr/local/nvidia/bin && ./nvidia-smi

Container:

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

If GPU information is returned, the device is available and the add-on has been installed.

Obtaining the Driver Link from Public Network

Log in to the CCE console.
Click Create Node and select the GPU node to be created in the Specifications area. The GPU card model of the node is displayed in the lower part of the page.

Visit https://www.nvidia.com/Download/Find.aspx?lang=en.
Select the driver information on the NVIDIA Driver Downloads page, as shown in Figure 1. Operating System must be Linux 64-bit.

Figure 1 Setting parameters
After confirming the driver information, click SEARCH. A page is displayed, showing the driver information, as shown in Figure 2. Click DOWNLOAD.

Figure 2 Driver information
Obtain the driver link in either of the following ways:
- Method 1: As shown in Figure 3, find url=/tesla/470.103.01/NVIDIA-Linux-x86_64-470.103.01.run in the browser address box. Then, supplement it to obtain the driver link https://us.download.nvidia.com/tesla/470.103.01/NVIDIA-Linux-x86_64-470.103.01.run. By using this method, you must bind an EIP to each GPU node.
- Method 2: As shown in Figure 3, click AGREE & DOWNLOAD to download the driver. Then, upload the driver to OBS and record the OBS URL. By using this method, you do not need to bind an EIP to GPU nodes.
  Figure 3 Obtaining the link

Obtaining the Driver Link from OBS

Upload the driver to OBS and set the driver file to public read.

When the node is restarted, the driver will be downloaded and installed again. Ensure that the OBS bucket link of the driver is valid.
In the bucket list, click a bucket name, and then the Overview page of the bucket is displayed.
In the navigation pane, choose Objects.
Select the name of the target object and copy the driver link on the object details page.