Elastic GPU Service:Automatically install or load Tesla drivers when creating a GPU-accelerated instance

Prerequisites

Before you begin, ensure that you have:

• A GPU-accelerated instance type that supports automatic driver installation (see the compatibility table below)

• Access to the Instance buy page

Compatibility

The following table lists supported instance types, public image versions, and corresponding driver versions.

ebmgn9g and ebmgn9ge do not support Ubuntu 22.04.

Steps

1. Go to the Instance buy page.

2. Select the Custom Launch tab.

3. Configure Billing Method, Region, Network and Zone, Instance Type, and Image. For details on all settings, see Configuration items.

   After the instance is created or started, the following applies to the Tesla driver:

   The automatic installation takes approximately 10 to 20 minutes. The actual time varies depending on the instance's internal network bandwidth and number of vCPUs. You can connect to the instance to monitor the installation progress. After the installation is complete, you can view the log file at /root/auto_install/auto_install.log. The following table describes the statuses displayed during installation.

   Installation state	Displayed information
   Installing	A progress bar is displayed.
   Installation successful	The message ALL INSTALL OK is displayed.
   Installation failed	The message INSTALL FAIL is displayed.

   Important

   The GPU is unavailable during the installation process. Do not perform any operations on the instance. If other GPU-related software is installed at the same time, the automatic installation may fail and make the instance unavailable.

4. In the Image section, on the Public Image tab, select a Linux operating system and version (for example, Alibaba Cloud Linux 3.2104 LTS 64-bit). Select the Auto-install GPU Driver checkbox, then choose the driver, CUDA, and cuDNN versions.

   

5. Follow the prompts to complete the purchase.

Monitor installation progress

Installation takes 10 to 20 minutes. The actual time varies depending on the instance's internal network bandwidth and vCPU count.

Connect to the instance to check progress. The log file is at /root/auto_install/auto_install.log.

Script parameters

The script installs the Tesla driver, CUDA, and cuDNN using .run packages. Set the three version variables to match your instance type and image. Refer to the compatibility table for supported version combinations.

The following example sets the Tesla driver to 550.127.08, CUDA to 12.4.1, and cuDNN to 9.2.0.82:

DRIVER_VERSION="550.127.08"
CUDA_VERSION="12.4.1"
CUDNN_VERSION="9.2.0.82"

Steps

1. Go to the Instance buy page.

2. Select the Custom Launch tab.

3. Configure Billing Method, Region, Network and Zone, Instance Type, and Image. For details on all settings, see Configuration items.

4. In the Advanced Settings (Optional) section, enter your installation script in the User Data field. The following is a sample script. Replace the version variables with values appropriate for your instance type and image.

   #!/bin/sh
   
   # Set the versions to install
   DRIVER_VERSION="550.127.08"
   CUDA_VERSION="12.4.1"
   CUDNN_VERSION="9.2.0.82"
   IS_INSTALL_eRDMA="FALSE"
   IS_INSTALL_RDMA="FALSE"
   INSTALL_DIR="/root/auto_install"
   
   # Download and run the installation script
   auto_install_script="auto_install_v4.0.sh"
   
   script_download_url=$(curl http://100.100.100.200/latest/meta-data/source-address | head -1)"/opsx/ecs/linux/binary/script/${auto_install_script}"
   echo $script_download_url
   
   rm -rf $INSTALL_DIR
   mkdir -p $INSTALL_DIR
   cd $INSTALL_DIR && wget -t 10 --timeout=10 $script_download_url && bash ${INSTALL_DIR}/${auto_install_script} $DRIVER_VERSION $CUDA_VERSION $CUDNN_VERSION $IS_INSTALL_RDMA $IS_INSTALL_eRDMA

   

5. Follow the prompts to complete the purchase.

To create a GPU-accelerated instance using the RunInstances API operation, pass the installation script in the UserData parameter. For details, see RunInstances.

If you did not install the driver during instance creation, run the script manually afterward: connect to the instance via SSH, create a file, paste the script content, and run it as a shell script. For connection methods, see Connection methods.

Modify the corresponding Tesla driver, CUDA, and cuDNN library versions based on the instance type and image that you created. For detailed information about supported versions, see the image versions and instance types supported by the Tesla driver when you select Public Image in Automatically install drivers by using a public image.

The following example shows how to set the Tesla driver version to 550.127.08, the CUDA version to 12.4.1, and the cuDNN version to 9.2.0.82:

DRIVER_VERSION="550.127.08"
CUDA_VERSION="12.4.1"
CUDNN_VERSION="9.2.0.82"