Enabling Trainers with GPUs in On-Premise Kubernetes Deployments

In recent versions of Hyperscience, we’ve made it possible to classify and extract data from Unstructured documents. However, automating these processes requires more computing resources than our automation capabilities for Structured and Semi-structured documents do. 

In order to process Unstructured documents in on-premise deployments of Hyperscience, you need to add a trainer that has both a GPU (graphics processing unit) and a CPU (central processing unit). GPUs have specialized cores that allow the system to perform multiple computations in parallel, reducing the time required to complete the complex operations required to train models for Unstructured documents. When you attach a trainer whose machine has a GPU, you can maximize the benefits of Unstructured Extraction. To learn more about this feature, see Field Identification. 

This article describes how to enable a trainer with both a GPU and a CPU in an on-premise Kubernetes deployment of Hyperscience. 

In order for our training to take advantage of the GPU on the on-premise Kubernetes cluster, we have to make sure that the Kubernetes nodes are configured with correct drivers or AMIs and that the containers can access the GPU.

We have covered several Kubernetes setup options below:

• Bare Metal

• AWS EKS On-prem

• Azure AKS On-prem

• Google GKE On-prem

Bare Metal K8S setup

1. Make sure your GPU hardware meets the requirements.

To process Unstructured documents in Hyperscience, your trainer needs an NVIDIA GPU. We used the NVIDIA Tesla T4 for our benchmarking. Other NVIDIA GPUs may perform slightly better or worse, depending on cores, RAM, and other factors.

For more information on the T4’s specifications, see TechPowerUp’s NVIDIA Tesla T4. To learn more about other NVIDIA GPUs, see TechPowerUp’s GPU Specs Database.

Machine sizing

We've completed benchmark tests for AWS's g4dn.4xlarge machine, and we recommend that machine or one of comparable size. For more details on this machine, see AWS's Amazon EC2 G4 Instances. 

2. Pre-installation Actions

a. Verify that your GPU supports CUDA.

CUDA is a parallel computing platform and programming model created by NVIDIA. Machine learning often uses CUDA-based libraries, SDKs, and other tools.

You can find out whether your GPU supports CUDA by running the following command:

lspci | grep -i nvidia

For more information, see NVIDIA’s CUDA GPUs - Compute Capability and NVIDIA CUDA Installation Guide for Linux.

b. Verify that you have a supported version of Linux.

Follow the instructions in NVIDIA’s NVIDIA CUDA Installation Guide for Linux to check your version of Linux. Then, make sure your Linux version is supported by the latest CUDA Toolkit by reviewing NVIDIA’s NVIDIA CUDA Toolkit Release Notes.  

c. Verify that the system has gcc installed.

The gcc compiler is required for development using the CUDA Toolkit. To make sure it is installed, follow the instructions in NVIDIA’s NVIDIA CUDA Installation Guide for Linux.

d. Verify that the system has the current Kernel headers and development packages installed.

Kernel headers are header files that specify the interface between the Linux kernel and userspace libraries and programs. The CUDA driver requires that the kernel headers and development packages for the running version of the kernel be installed at the time of the driver installation, as well whenever the driver is rebuilt. For example, if your system is running kernel version 3.17.4-301, the 3.17.4-301 kernel headers and development packages must also be installed.

To verify that these requirements are met, run the following command:

apt-get install linux-headers-$(uname -r)

For more information and commands for various Linux distributions, see NVIDIA’s NVIDIA CUDA Installation Guide for Linux.

3. Install the CUDA drivers.

Follow these steps to make sure that you have the most current and correct CUDA drivers installed.

a. Remove any CUDA drivers installed on the system.

Compatibility between the CUDA Toolkit and CUDA drivers is crucial. In our Docker-based system, the CUDA Toolkit is installed in the trainer images, and you need to make sure that they match the CUDA drivers that are installed in the host OS. The current CUDA Toolkit version installed should be compatible with the latest available CUDA driver. For details on toolkit and driver versions, see NVIDIA’s NVIDIA CUDA Toolkit Release Notes.

An example command for removing these drivers appears below:

apt-get clean; apt-get update; apt-get purge -y cuda*; apt-get purge -y nvidia-*; apt-get -y autoremove

You can tailor this command to match your Linux distribution.

b. Install the NVIDIA Container Toolkit.

The Docker host needs to be prepared before it can expose your GPU hardware to the containers. Although containers share your host’s kernel, they cannot access information on the system packages you have installed. A plain container will lack the device drivers that interface with your GPU. You can activate support for NVIDIA GPUs by installing NVIDIA’s Docker Container Toolkit:

distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
      && curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
      && curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.list | \
            sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
            tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
      apt-get update && apt-get install -y nvidia-container-toolkit

Configure the container runtime by running the nvidia-ctk command shown below.

The nvidia-ctk command modifies the /etc/docker/daemon.json file on the host. The file is updated so that Docker can use the NVIDIA Container Runtime.

nvidia-ctk runtime configure --runtime=docker
systemctl restart docker

Inspect your /etc/docker/daemon.json file to confirm that the configured container runtime has been changed. The NVIDIA Toolkit will handle the injection of GPU device connections when new containers start.

c. Install the latest CUDA drivers.

Running the following command installs the latest CUDA driver versions, which should be compatible with the container toolkit:

distribution=$(. /etc/os-release;echo $ID$VERSION_ID | sed -e 's/\.//g')
     wget https://developer.download.nvidia.com/compute/cuda/repos/$distribution/x86_64/cuda-keyring_1.0-1_all.deb
     dpkg -i cuda-keyring_1.0-1_all.deb
     apt-get update && apt-get -y install cuda-drivers

d. Deploy the nvidia-device-plugin DaemonSet. 

To deploy the DaemonSet, follow the instructions in the “Enabling GPU Support in Kubernetes” section of the NVIDIA device plugin for Kubernetes documentation in GitHub. 

4. Change the model type for the layout you want to use with Unstructured Extraction.

Before you can use your GPU-based trainer, you first need to determine which layout you’d like to use Unstructured Extraction with. Then, you need to change the model type for that layout to UNSTRUCTURED_EXTRACTION. The system will use the GPU for Unstructured Extraction of data from that layout’s documents.

To do so:

1. Go to /admin/form_extraction/template/.

2. Find the record for the layout you’d like to use Unstructured Extraction with, and click on its UUID.

3. In the Flex engine type for training drop-down list, select UNSTRUCTURED_EXTRACTION.

4. Click Save.

AWS EKS On-prem setup

1. You can use the pre-built GPU AMIs made available by AWS with all the necessary drivers installed. For more information, see the “Amazon EKS optimized accelerated Amazon Linux AMIs” section of AWS’s Amazon EKA optimized Amazon Linux AMIs. 

   • For a list of pre-built GPU AMIs based on Kubernetes version, see Amazon EKS AMI’s Releases list in GitHub.

2. Deploy nvidia-plugin-daemonset by following the instructions in the “Enabling GPU Support in Kubernetes” section of the NVIDIA device plugin for Kubernetes documentation in GitHub.

Azure AKS On-prem setup

To enable GPU for Azure AKS, follow the steps in Microsoft’s Use GPUs for compute-intensive workloads on Azure Kubernetes Service (AKS). Doing so sets up the GPU nodes and installs the nvidia-device-plugin DaemonSet.

Google GKE On-prem setup

Follow Google Cloud’s Run GPUs in GKE Standard node pools to set up GPU node pools with drivers and the DaemonSet.