How do I use a graphics processing unit (GPU) with Dataiku DSS?

Prerequisites

• A Linux server with an NVIDIA GPU. These instructions were written using CentOS 7; Ubuntu 14.04 or later is known to work as well, though installation instructions may differ.

• Dataiku DSS installed

Libraries version

We aim to install the following Python libraries:

• Keras 1.2.2 (version required in fast.ai)

• Tensorflow-gpu 1.4.1

To do so, we must install:

• CUDA 8.0

• cuDNN 6.0

Install NVIDIA driver and the CUDA and cuDNN libraries

Install NVIDIA driver

Before installing the driver we need to install required libraries, using the yum package-management utility shipped with CentOS:

sudo yum update
sudo yum install kernel-devel kernel-headers gcc gcc-c++ make wget epel-release dkms libstdc++.i686 bzip2 python-pip
sudo reboot

The new kernel should now be loaded. You can verify that your kernel is the same as the installed source by comparing the results of the following two commands:

uname -r
rpm -q kernel-devel

If the version numbers are not the same, you can upgrade and reboot:

sudo yum -y upgrade kernel kernel-devel
sudo reboot

By default, CentOS comes with a "nouveau" driver for the GPU. You can verify it with the command

lsmod | grep nouv

We need to blacklist this driver in order to let the GPU use NVIDIA. Hence, you need to create a file at: /etc/modprobe.d/blacklist-nouveau.conf, containing:

blacklist nouveau
options nouveau modeset=0

To enforce the blacklist, run the commands

sudo dracut --force  
sudo reboot

You can verify that the driver is not used anymore by re-running the command lsmod | grep nouv, which should not display anything this time.

We can now install the NVIDIA driver with its official runfile. You need to download the appropriate runfile from the NVIDIA website. Be careful to take NVIDIA for CUDA 8.0, which is the last supported version of CUDA for Tensorflow.

For our instance, it was the following parameters:

...and can be downloaded with the following command:

wget http://us.download.nvidia.com/XFree86/Linux-x86_64/384.66/NVIDIA-Linux-x86_64-384.66.run

First make the file executable (the filename may be slightly different according to the version you have downloaded):

sudo chmod +x NVIDIA-Linux-x86_64-384.66.run

Then execute it:

./NVIDIA-Linux-x86_64-384.66.run

Follow the instructions in the command line installation interface, using default options.

NVIDIA drivers are now installed! You can check that it worked by running the command nvidia-smi. That should display something like:

Install the CUDA library

First you need to download the CUDA toolkit. You want to install CUDA 8.0, and make the appropriate selections for your platform:

Select the runfile installer. For our instance, the link is:

wget https://developer.nvidia.com/compute/cuda/8.0/Prod2/local_installers/cuda_8.0.61_375.26_linux-run

As before, you need to make the file executable:

sudo chmod +x cuda_8.0.61_375.26_linux-run

Then run this command to retrieve the installer:

sudo ./cuda_8.0.61_375.26_linux-run --extract=$HOME

Next, find the installer (it starts with "cuda-linux64-rel") and run it:

sudo ./cuda-linux64-rel-8.0.61-21551265.run

Then follow the installation steps, using default options. Remember to say yes to the question "Would you like to add desktop menu shortcuts?".

Finally, you need to add CUDA to the PATH and create LD_LIBRARY_PATH. To do so, add the following line to your ~/.bashrc:

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

Be sure to run source ~/.bashrc afterward, or to reboot your instance so that your changes are taken into account.

You can check that CUDA is installed by running the command nvcc --version.

Installing the cuDNN library

First register as a member of the NVIDIA developer program.

Then download cuDNN 6.0 for CUDA 8.0 for Linux:

Then, transfer it to your instance, for instance using:

scp -i <MY_PRIVATE_KEY_FILE> <PATH_TO_CUDNN_ARCHIVE> centos@<MY_PUBLIC_DNS_IPV4>:~

Decompress the file:

tar -xzf cudnn-8.0-linux-x64-v6.0.tgz

Then copy/paste the library to your CUDA installation directory. Go to the CUDA folder created at the previous step and run the commands:

cd cuda
cp include/cudnn.h /usr/local/cuda/include/
cp lib64/libcudnn* /usr/local/cuda/lib64/

cuDNN is now installed!

Set Up A Deep Learning Code Environment in Dataiku DSS

If you plan to use the Deep Learning for Images plugin, follow the instructions in the Howto for installing the plugin and it will set up its own code environment.

To create a custom code environment, follow the reference documentation to create a Python environment and install the following libraries (note the following copy and paste-able text assumes you are using conda):

# pip
tensorflow-gpu==1.4.1
keras==1.2.2
matplotlib
Pillow
scikit-learn
scipy
bcolz
h5py

# conda
libgcc
mkl-service

Now you can use your environment from a notebook in your own DSS project.

You can test tensorflow-gpu from your notebook:

from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())

It should give you a message like this:

If any GPU is listed, it means that you can use it with tensorflow. Congratulations!