Nvidia Docker Ubuntu 16.04



  1. Also, you need to install the Docker Engine Utility for NVIDIA GPUs, NVIDIA Docker. To access the Azure RDMA network, use an RDMA-capable VM size. Necessary RDMA drivers are installed in the CentOS HPC and Ubuntu images supported by Batch.
  2. Certificaat Thuiswinkel.org verklaart dat haar lid: het Certificaat Thuiswinkel Waarborg mag voeren. Dit betekent dat Zalando.nl als webshop is gecertificeerd door de Stichting Certificering Thuiswinkel Waarborg.
-->

Read about the latest tech news and developments from our team of experts, who provide updates on the new gadgets, tech products & services on the horizon.

Azure Batch lets you run and scale large numbers of batch computing jobs on Azure. Batch tasks can run directly on virtual machines (nodes) in a Batch pool, but you can also set up a Batch pool to run tasks in Docker-compatible containers on the nodes. This article shows you how to create a pool of compute nodes that support running container tasks, and then run container tasks on the pool.

The code examples here use the Batch .NET and Python SDKs. You can also use other Batch SDKs and tools, including the Azure portal, to create container-enabled Batch pools and to run container tasks.

Why use containers?

Using containers provides an easy way to run Batch tasks without having to manage an environment and dependencies to run applications. Containers deploy applications as lightweight, portable, self-sufficient units that can run in several different environments. For example, build and test a container locally, then upload the container image to a registry in Azure or elsewhere. The container deployment model ensures that the runtime environment of your application is always correctly installed and configured wherever you host the application. Container-based tasks in Batch can also take advantage of features of non-container tasks, including application packages and management of resource files and output files.

Prerequisites

You should be familiar with container concepts and how to create a Batch pool and job.

  • SDK versions: The Batch SDKs support container images as of the following versions:

    • Batch REST API version 2017-09-01.6.0
    • Batch .NET SDK version 8.0.0
    • Batch Python SDK version 4.0
    • Batch Java SDK version 3.0
    • Batch Node.js SDK version 3.0

  • Accounts: In your Azure subscription, you need to create a Batch account and optionally an Azure Storage account.

  • A supported VM image: Containers are only supported in pools created with the Virtual Machine Configuration, from images detailed in the following section, 'Supported virtual machine images.' If you provide a custom image, see the considerations in the following section and the requirements in Use a managed custom image to create a pool of virtual machines.

Keep in mind the following limitations:

  • Batch provides RDMA support only for containers running on Linux pools.

  • For Windows container workloads, we recommend choosing a multicore VM size for your pool.

Supported virtual machine images

Use one of the following supported Windows or Linux images to create a pool of VM compute nodes for container workloads. For more information about Marketplace images that are compatible with Batch, see List of virtual machine images.

Windows support

Batch supports Windows server images that have container support designations. Typically these image sku names are suffixed with -with-containers or -with-containers-smalldisk. Additionally, the API to list all supported images in Batch will denote a DockerCompatible capability if the image supports Docker containers.

Python

You can also create custom images from VMs running Docker on Windows.

Linux support

For Linux container workloads, Batch currently supports the following Linux images published by Microsoft Azure Batch in the Azure Marketplace without the need for a custom image.

VM sizes without RDMA

  • Publisher: microsoft-azure-batch
    • Offer: centos-container
    • Offer: ubuntu-server-container

VM sizes with RDMA

  • Publisher: microsoft-azure-batch
    • Offer: centos-container-rdma
    • Offer: ubuntu-server-container-rdma

These images are only supported for use in Azure Batch pools and are geared for Docker container execution. They feature:

  • A pre-installed Docker-compatible Moby container runtime

  • Pre-installed NVIDIA GPU drivers and NVIDIA container runtime, to streamline deployment on Azure N-series VMs

  • Pre-installed/pre-configured image with support for Infiniband RDMA VM sizes for images with the suffix of -rdma. Currently these images do not support SR-IOV IB/RDMA VM sizes.

You can also create custom images from VMs running Docker on one of the Linux distributions that is compatible with Batch. If you choose to provide your own custom Linux image, see the instructions in Use a managed custom image to create a pool of virtual machines.

For Docker support on a custom image, install Docker Community Edition (CE) or Docker Enterprise Edition (EE).

Additional considerations for using a custom Linux image:

  • To take advantage of the GPU performance of Azure N-series sizes when using a custom image, pre-install NVIDIA drivers. Also, you need to install the Docker Engine Utility for NVIDIA GPUs, NVIDIA Docker.

  • To access the Azure RDMA network, use an RDMA-capable VM size. Necessary RDMA drivers are installed in the CentOS HPC and Ubuntu images supported by Batch. Additional configuration may be needed to run MPI workloads. See Use RDMA-capable or GPU-enabled instances in Batch pool.

Container configuration for Batch pool

To enable a Batch pool to run container workloads, you must specify ContainerConfiguration settings in the pool's VirtualMachineConfiguration object. (This article provides links to the Batch .NET API reference. Corresponding settings are in the Batch Python API.)

You can create a container-enabled pool with or without prefetched container images, as shown in the following examples. The pull (or prefetch) process lets you pre-load container images from either Docker Hub or another container registry on the Internet. For best performance, use an Azure container registry in the same region as the Batch account.

The advantage of prefetching container images is that when tasks first start running they don't have to wait for the container image to download. The container configuration pulls container images to the VMs when the pool is created. Tasks that run on the pool can then reference the list of container images and container run options.

Pool without prefetched container images

To configure a container-enabled pool without prefetched container images, define ContainerConfiguration and VirtualMachineConfiguration objects as shown in the following examples. These examples use the Ubuntu Server for Azure Batch container pools image from the Marketplace.

Prefetch images for container configuration

To prefetch container images on the pool, add the list of container images (container_image_names, in Python) to the ContainerConfiguration.

The following basic Python example shows how to prefetch a standard Ubuntu container image from Docker Hub.

The following C# example assumes that you want to prefetch a TensorFlow image from Docker Hub. This example includes a start task that runs in the VM host on the pool nodes. You might run a start task in the host, for example, to mount a file server that can be accessed from the containers.

Prefetch images from a private container registry

You can also prefetch container images by authenticating to a private container registry server. In the following examples, the ContainerConfiguration and VirtualMachineConfiguration objects prefetch a private TensorFlow image from a private Azure container registry. The image reference is the same as in the previous example.

Nvidia Docker Ubuntu 16.04 Debian

Container settings for the task

To run a container task on a container-enabled pool, specify container-specific settings. Settings include the image to use, registry, and container run options.

  • Use the ContainerSettings property of the task classes to configure container-specific settings. These settings are defined by the TaskContainerSettings class. Note that the --rm container option doesn't require an additional --runtime option since it is taken care of by Batch.

  • If you run tasks on container images, the cloud task and job manager task require container settings. However, the start task, job preparation task, and job release task do not require container settings (that is, they can run within a container context or directly on the node).

  • For Windows, tasks must be run with ElevationLevel set to admin.

  • For Linux, Batch will map the user/group permission to the container. If access to any folder within the container requires Administrator permission, you may need to run the task as pool scope with admin elevation level. This will ensure Batch runs the task as root in the container context. Otherwise, a non-admin user may not have access to those folders.

  • For container pools with GPU-enabled hardware, Batch will automatically enable GPU for container tasks, so you shouldn't include the –gpus argument.

Container task command line

When you run a container task, Batch automatically uses the docker create command to create a container using the image specified in the task. Batch then controls task execution in the container.

As with non-container Batch tasks, you set a command line for a container task. Because Batch automatically creates the container, the command line only specifies the command or commands that will run in the container.

If the container image for a Batch task is configured with an ENTRYPOINT script, you can set your command line to either use the default ENTRYPOINT or override it:

  • To use the default ENTRYPOINT of the container image, set the task command line to the empty string '.

  • To override the default ENTRYPOINT, or if the image doesn't have an ENTRYPOINT, set a command line appropriate for the container, for example, /app/myapp or /bin/sh -c python myscript.py.

Optional ContainerRunOptions are additional arguments you provide to the docker create command that Batch uses to create and run the container. For example, to set a working directory for the container, set the --workdir <directory> option. See the docker create reference for additional options.

Container task working directory

A Batch container task executes in a working directory in the container that is very similar to the directory Batch sets up for a regular (non-container) task. Note that this working directory is different from the WORKDIR if configured in the image, or the default container working directory (C: on a Windows container, or / on a Linux container).

For a Batch container task:

  • All directories recursively below the AZ_BATCH_NODE_ROOT_DIR on the host node (the root of Azure Batch directories) are mapped into the container
  • All task environment variables are mapped into the container
  • The task working directory AZ_BATCH_TASK_WORKING_DIR on the node is set the same as for a regular task and mapped into the container.

These mappings allow you to work with container tasks in much the same way as non-container tasks. For example, install applications using application packages, access resource files from Azure Storage, use task environment settings, and persist task output files after the container stops.

Troubleshoot container tasks

If your container task doesn't run as expected, you might need to get information about the WORKDIR or ENTRYPOINT configuration of the container image. To see the configuration, run the docker image inspect command.

If needed, adjust the settings of the container task based on the image:

  • Specify an absolute path in the task command line. If the image's default ENTRYPOINT is used for the task command line, ensure that an absolute path is set.
  • In the task's container run options, change the working directory to match the WORKDIR in the image. For example, set --workdir /app.

Container task examples

The following Python snippet shows a basic command line running in a container created from a fictitious image pulled from Docker Hub. Here, the --rm container option removes the container after the task finishes, and the --workdir option sets a working directory. The command line overrides the container ENTRYPOINT with a simple shell command that writes a small file to the task working directory on the host.

The following C# example shows basic container settings for a cloud task:

Next steps

  • For easy deployment of container workloads on Azure Batch through Shipyard recipes, see the Batch Shipyard toolkit .
  • For information on installing and using Docker CE on Linux, see the Docker documentation.
  • Learn how to Use a managed custom image to create a pool of virtual machines.
  • Learn more about the Moby project, a framework for creating container-based systems.

At Build 2020 Microsoft announced support for GPU compute on Windows Subsystem for Linux 2. Ubuntu is the leading Linux distribution for WSL and a sponsor of WSLConf. Canonical, the publisher of Ubuntu, provides enterprise support for Ubuntu on WSL through Ubuntu Advantage.

This guide will walk early adopters through the steps on turning their Windows 10 devices into a CUDA development workstation with Ubuntu on WSL.

For our purposes we will be setting up Jupyter Notebook in Docker with CUDA on WSL. These instructions can be adapted to set up other CUDA GPU compute workloads on WSL.

Install Windows 10 Insiders Dev Channel

To begin, we need the latest Windows 10 Insider build released today, June 17, 2020. You will need to register as a Windows Insider if you have not already, enroll your device in the Dev Channel (previously known as the ‘Fast Ring’), and then upgrade to Windows 10 build number 20150 or above.

Consult the Windows Insider documentation for more information on registering as an Insider, enrolling your device, and upgrading your machine to the Dev Channel.

Enable WSL 2

In future updates to Windows you will simply need to use the following to enable WSL:

For now, open PowerShell as Administrator.

First enable WSL 1:

Then enable WSL 2:

Automator mac remove. Restart Windows 10:

This step will become redundant in the future after WSL 2 becomes the default, but for now return to PowerShell and make WSL 2 your default before installing Ubuntu:

To read more about Ubuntu on WSL, visit ubuntu.com/wsl. For a more detailed look at enabling WSL on Windows, check out our tutorial. To convert existing WSL 1 installs to WSL 2, see my blog on the general availability of WSL 2.

Install Ubuntu on WSL

Install Ubuntu from the Microsoft Store:

For other ways to install Ubuntu on WSL, see our WSL wiki.

Install Windows Terminal

Optionally, you may install the new Windows Terminal from the Microsoft Store. It is has many features, such as GPU acceleration and customizability, that improves the Ubuntu experience on WSL over the traditional Windows console.

Windows Terminal can also be installed from the project’s GitHub page.

Setup Ubuntu on WSL

Open Ubuntu from the Windows Start Menu and configure your WSL user. This user is separate from your Windows user:

If you downloaded Windows Terminal you can then close the old console and re-open Ubuntu from the drop-down options in the new Terminal:

Now check to make sure you are running the correct WSL 2 Linux kernel.

In Ubuntu:

You will need kernel 4.19.121 or higher. If you do not have kernel 4.19.121, first try:

If that does not work, make sure you have ‘Receive updates for other Microsoft products when you update Windows’ checked:

And then re-run Windows updates:

Install Nvidia Drivers on Windows 10

Next, download the appropriate driver for your GeForce or Quadro Nvidia card.

In the next few months, the NVIDIA driver will be distributed via Windows Update, which will manually downloading and installing the driver unnecessary.

You will need to join the Nvidia Developer Program for early access to the driver for now. You can read more about CUDA on WSL on the Nvidia Developer blog.

Install Docker in WSL

Install Nvidia Container Toolkit

Set the distribution variable, import the Nvidia repository GPG key, and then add the Nvidia repositories to the Ubuntu apt package manager:

Refresh the Ubuntu apt repositories and then install the Nvidia runtime:

Close all open terminals of Ubuntu, open a PowerShell terminal, and manually shutdown Ubuntu:

Ubuntu 16.04 Nvidia-docker 설치

Test GPU Compute

Open a new Ubuntu terminal and start Docker:

Nvidia Docker Ubuntu 16.04

Nvidia Docker Ubuntu 16.04 Installer

Nvidia docker ubuntu 16.04 centos

And then run:

If everything is configured correctly, the output should resemble:

Start A TensorFlow Container

Open a new Ubuntu terminal and run:

Open a second Ubuntu terminal, type wslview, copy and paste the notebook URL, but then edit the URL from 127.0.0.1 to localhost:

Your default browser on Windows will launch with a GPU-accelerated Jupyter notebook:

You are now all set to begin using TensorFlow with CUDA on Ubuntu WSL.

Related blog posts

  • Ubuntu: Ubuntu on WSL 2 Is Generally Available
  • Ubuntu: New GPU and GUI features announced for WSL at Build
  • Microsoft: GPU Compute, WSL Install and WSL Update arrive in the latest Insider build for the Windows Subsystem for Linux
  • Nvidia: Announcing CUDA on Windows Subsystem for Linux 2

Additional resources

Talk to us today

Interested in running Ubuntu in your organisation?

Newsletter signup