Although GPUs available with the current generation of AWS g2.2xlarge/8xlarge instance types might not be the fastest in the market, using AWS for GPU computing offers the flexibility of training multiple models simultaneously at minimal upfront cost.
Step 1: Create a g2.2xlarge instance using Ubuntu 14.04 LTS.
Right now it costs around $0.65 per hour for this instance type. Creating spot instances might offer additional cost savings, with the added caveat that the instance might be reclaimed when demand rises. If you plan on using Spot instances, be sure to check out the pricing history for this instance type/region to ensure that your bid price is appropriately set.
Step 2: Install essentials
sudo apt-get update && sudo apt-get upgrade sudo apt-get install build-essential sudo apt-get install git vim python-pip python3-pip liblapack-dev cython cython3 gfortran
Step 3: Download cuda
The package we download below includes both the nvidia driver and cuda. However, the nvidia driver version provided in this package is not the latest (aws gpu users require a newer version of the driver that we will download in step -6). We will use the cuda version provided in this package though.
mkdir -p install/cuda cd install/cuda wget http://developer.download.nvidia.com/compute/cuda/7.5/Prod/local_installers/cuda_7.5.18_linux.run chmod +x cuda_7.5.18_linux.run mkdir cuda_installer ./cuda_7.5.18_linux.run -extract=`pwd`/cuda_installer
Step 4: Install additional drivers that are not present in the base kernel package
sudo apt-get install linux-image-extra-virtual (choose "use local version" if asked about grub version)
Step 5: Disable nouveau since it conflicts with NVIDIA's kernel module
Nouveau is an open-source driver for Nvidia graphics cards. In step-6, we'll be installing the driver provided by nvidia. So we disable the default driver.
$ sudo vi /etc/modprobe.d/blacklist-nouveau.conf Add the below ---- blacklist nouveau blacklist lbm-nouveau options nouveau modeset=0 alias nouveau off alias lbm-nouveau off ---- echo options nouveau modeset=0 | sudo tee -a /etc/modprobe.d/nouveau-kms.conf sudo update-initramfs -u sudo reboot * After reboot sudo apt-get install linux-source sudo apt-get install linux-headers-`uname -r`
Step 6: Download and install the NVIDIA driver
NOTE: This uses the newer NVIDIA driver 352.63 that has a fix required for AWS users
cd ~/install/cuda
mkdir cuda_installer-352.63-aws
cd cuda_installer-352.63-aws
wget http://us.download.nvidia.com/XFree86/Linux-x86_64/352.63/NVIDIA-Linux-x86_64-352.63.run
chmod +x NVIDIA-Linux-x86_64-352.63.run
sudo ./NVIDIA-Linux-x86_64-352.63.run
Step 7: Verify nvidia driver installation
Run nvidia-smi to view available GPUs
$ nvidia-smi
Tue Jan 19 06:58:00 2016
+------------------------------------------------------+
| NVIDIA-SMI 352.63 Driver Version: 352.63 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GRID K520 Off | 0000:00:03.0 Off | N/A |
| N/A 34C P0 36W / 125W | 11MiB / 4095MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
Step 8: Install CUDA
sudo ./cuda-linux64-rel-7.5.18-19867135.run
Default install path: /usr/local/cuda-7.5
Symbolic link created: /usr/local/cuda -> /usr/local/cuda-7.5
Step 9: Install CUDA samples
sudo ./cuda-samples-linux-7.5.18-19867135.run Default install path: /usr/local/cuda-7.5/samples
Step 10: In .bashrc, add the following lines
export PATH=$PATH:/usr/local/cuda-7.5/bin export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-7.5/lib64
Remember to source the bashrc file
source ~/.bashrc
Step 11: Install python packages
You could install a distribution like anaconda, or alternatively install select packages as done below.
sudo pip3 install numpy sudo pip3 install scipy sudo apt-get install libyaml-dev sudo pip3 install --upgrade pyyaml sudo pip3 install --upgrade six
Step 12: Install and configure theano
sudo pip3 install --upgrade --no-deps git+git://github.com/Theano/Theano.git vi ~/.theanorc ---- [global] floatX = float32 device = gpu optimizer = fast_run [lib] cnmem = 0.9 [nvcc] fastmath = True [blas] ldflags = -llapack -lblas ----
Step 13: Install keras
cd ~/install git clone https://github.com/fchollet/keras.git keras cd keras sudo python3 setup.py install
Step 14: Verify
Run the MNIST CNN example to verify that the installation succeeded
cd keras/examples THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python3 mnist_cnn.py
Note that you can monitor GPU usage by using the nvidia-smi command (prints status every sec)
nvidia-smi -l 1
NOTE: You could also install cuDNN for additional speed up, although you would need to apply for the nvidia developer program in order to download it.
