Red Cloud GPU Image Usage

This wiki provides documentation for creating and using a GPU instance in Red Cloud from a Ubuntu-based (gpu-accelerated-ubuntu-2020-08 or a CentOS-based (gpu-accelerated-centos-2020-08 GPU image. The image features:

GPU acceleration via CUDA 10.1.
Anaconda distribution to facilitate usage of platforms such as TensorFlow,
Docker-containerized Jupyter Notebook servers, and
MATLAB R2019a.

A test application using the Python neural network library Keras, which runs on top of the TensorFlow framework, is provided to check and test GPU utilization.

The intent is to help you to get a quick start on using GPU-accelerated software in your work, with minimal effort on your part.

Create a Server Instance

First you'll want to take care of a few preliminaries covered elsewhere in this wiki:

Now you're ready to launch a new server instance:

In the OpenStack web interface Dashboard, navigate to Compute > Instances.
.
Click on “Launch Instance”.
Make selections on the Launch Instance screen:
- Details tab: enter the instance name (your choice).
- Source tab:
  - Boot Source = “Image”.
  - Among the list of images, choose “gpu-accelerated-centos-2020-08” for Centos 7.8 or “gpu-accelerated-ubuntu-2020-08” for Ubuntu 18.04 LTS by clicking on the ⬆️ button.
  - Volume Size: increase as necessary. For reference, the Anaconda distribution alone is 19GB.
- Flavor tab: select instance server configurations. Choose between "c4.t1.m20" (Tesla T4 GPU) and "c14.g1.m60" (Tesla V100 GPU for large jobs). If you need multiple GPUs on one instance, please reach out for special accommodation.
- Networks tab: “public” is a good default.
- Security Groups tab: Click ⬇️ on the “default” group to deselect the default security group. Click ⬆️ to select the custom security group you created above.
- Key Pair tab: Select the ssh key pair you had created above.
- Other tabs do not require modification.
Click the "Launch Instance" button to create your server instance.
Note the IP Address of your newly-created instance.

In a new terminal window, Chrome extension ssh or Putty, SSH into the instance:

For Ubuntu instance:

ssh -i ~/.ssh/id_rsa ubuntu@<IP_Address_from_previous_step>

For CentOS instance:

ssh -i ~/.ssh/id_rsa centos@<IP_Address_from_previous_step>

This assumes that your matching private key is saved as ~/.ssh/id_rsa.

CAC Recommends

Monthly security update patching:

For Ubuntu instance:
```
sudo apt dist-upgrade
```
For CentOS instance:
```
sudo yum update
```

Installed Software

CUDA

Check version with yum info cuda (not applicable for Ubuntu)

Check CUDA version currently in use with nvcc --version

(base) ubuntu@shl1-test1:~$ nvcc --version
    nvcc: NVIDIA (R) Cuda compiler driver
    Copyright (c) 2005-2019 NVIDIA Corporation
    Built on Sun_Jul_28_19:07:16_PDT_2019
    Cuda compilation tools, release 10.1, V10.1.243

NVIDIA Driver

Check version with dkms status<
- Run sudo yum install dkms (CentOS) or sudo apt install dkms if “command not found”

Check detection of GPU devices by CUDA via NVIDIA’s drivers: nvidia-smi

(base) ubuntu@shl1-test1:~$ nvidia-smi
    Sat Aug 29 01:38:55 2020     

    +-----------------------------------------------------------------------------+
    | NVIDIA-SMI 450.57       Driver Version: 450.57       CUDA Version: 11.0     |
    |-------------------------------+----------------------+----------------------+
    | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
    | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
    |                               |                      |               MIG M. |
    |===============================+======================+======================|
    |   0  Tesla T4            Off  | 00000000:00:05.0 Off |                    0 |
    | N/A   41C    P0    27W /  70W |      0MiB / 15109MiB |      6%      Default |
    |                               |                      |                  N/A |
    +-------------------------------+----------------------+----------------------+                                                                             

    +-----------------------------------------------------------------------------+
    | Processes:                                                                  |
    |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
    |        ID   ID                                                   Usage      |
    |=============================================================================|
    |  No running processes found                                                 |
    +-----------------------------------------------------------------------------+

Note that the CUDA version displayed on top right of this output is not necessarily the CUDA version currently in use.

Anaconda

Check overarching Anaconda version with conda list anaconda$
Check “conda” package manager version with conda -V
Check list of packages installed with conda list
If desired, update to latest version with conda update --all
For environment information, such as base paths or Python platform versioning, use conda info
An Anaconda environment should already be activated upon startup, indicated by base preceding each command line prompt.
```
(base) ubuntu@shl1-test1:~$
```

This “base” environment applies specific settings to your shell for the choice of a Python interpreter and its associated modules and libraries. Use which python for identifying the environment’s Python location; python -V for Python version.

Docker

To start docker using docker:

sudo usermod -aG docker $(whoami)
sudo service docker start

To view existing images pulled onto this instance,
```
docker images
```
To view existing containers, both running and exited,
```
docker ps -a
```

To enter terminal of a running container,

docker exec -it <Container Name or Container ID> bash

To restart an exited container (starts from latest first),
```
docker start `docker ps -q -l`
```
To exit container Terminal without exiting or killing the container, press CTRL+p then CTRL+q.

For more information on Docker, see the official Docker documentation or the Aristotle user documentation.

MATLAB

MATLAB is installed in /usr/local/bin/matlab. See GPU Computing in MATLAB page from MathWorks for more information.

Testing with Sample Application with Virtualization Only

Each of these frameworks can be set up in their own Conda environments.

Use conda env list to see the list of existing usable environments.

(base) ubuntu@shl1-test1:~$ conda env list
# conda environments:
#
base                  *  /home/ubuntu/anaconda3
pytorch                  /home/ubuntu/anaconda3/envs/pytorch
tf_gpu                   /home/ubuntu/anaconda3/envs/tf_gpu

Use conda deactivate to exit current environment and return to “base” environment.

Keras-GPU on TensorFlow-GPU

Activate the Virtual Environment for TensorFlow
- With the “base” Anaconda environment still activated,
```
conda create -n tf_gpu tensorflow-gpu
conda activate tf_gpu
```
Install necessary packages:
```
conda install tensorflow-gpu keras-gpu
```
Ensure the iPython command-line terminal used is from within the environment, not an external version:
```
conda install -c anaconda ipython
```
which ipython should output ~/anaconda3/envs/tf_gpu/bin/ipython

Start ipython and paste this sample app at the ipython prompt:

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
tf.config.list_physical_devices('GPU')

You should get output like this:

Out[3]: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

PyTorch

Activate the Virtual Environment for PyTorch
- With the “base” Anaconda environment still activated,
```
conda create -n pytorch
conda activate pytorch
```

Install necessary packages:

conda install pytorch torchvision -c pytorch
pip install pycuda

Ensure the iPython command-line terminal used is from within the environment, not an external version:
```
conda install -c anaconda ipython
```

Start ipython and paste this sample app at the ipython prompt:

import pycuda 
import pycuda.driver as drv  
drv.init() 
print('CUDA device query (PyCUDA version) \n') 
print('Detected {} CUDA Capable device(s) \n'.format(drv.Device.count())) 
for i in range(drv.Device.count()): 
     gpu_device = drv.Device(i) 
     print('Device {}: {}'.format( i, gpu_device.name() ) ) 
     compute_capability = float( '%d.%d' % gpu_device.compute_capability() )  
     print('\t Compute Capability:{}'.format(compute_capability))  
     print('\t Total Memory: {} megabytes'.format(gpu_device.total_memory()//(1024**2 )))

You should get the output

CUDA device query (PyCUDA version) 
Detected 1 CUDA Capable device(s) 
Device 0: Tesla T4
     Compute Capability:7.5
     Total Memory: 15109 megabytes

CUDA device query (PyCUDA version) 
Detected 1 CUDA Capable device(s) 
Device 0: Tesla V100-PCIE-16GB
     Compute Capability:7.0
     Total Memory: 16160 megabytes

Testing with Sample Application: with Virtualization + Docker Containerization

Keras GPU on TensorFlow-GPU

Create a new container:
```
docker run --gpus all -v $(realpath ~/notebooks):/tf/notebooks -p 8000:8000 tensorflow/tensorflow:latest-gpu-jupyter sleep 100000 &
```
- The --gpus all tag is for employing devices as detected by nvidia-smi.
- tensorflow/tensorflow:latest-gpu-jupyter is the specific image used.
- Read more about this image and other possible tags at DockerHub.
- The container runs a Jupyter Notebook server, exposed on host port 8888, from your own directory (here stated as ~/notebooks).
Test run the new container, then exit and disconnect:
- Use docker ps to check the Container ID (first column) or Name (last column)
- Run docker exec -it <Container ID or Name> bash
- Press CTRL+p and CTRL+q to exit container without killing it.
- Press CTRL+d to end SSH connection to VM.
Re-connect to the VM via SSH Tunneling, then restart the Docker TensorFlow Container: ssh -L 8000:localhost:8000 centos@<IP> OR ssh -L 8000:localhost:8000 ubuntu@<IP>

Then

docker exec -it <Container ID or Name> bash

Spin up a Jupyter Notebook within this Docker container:
```
jupyter notebook --ip 0.0.0.0 --port 8000 --allow-root
```
Copy the token provided above for use in the next step. Note: “No web browser found” error may occur, ignore if the next step is successful. If not, Ctrl+c to stop this current server, and try again.
Navigate to http://localhost:8000/ on a local browser. Enter the token from the previous step in the requested field.
Navigate to notebooks, then create a new Python3 Notebook or use an existing one.

Paste in and run this Keras sample application:

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
tf.config.list_physical_devices('GPU')

You should get the following output:

    [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search