Red Cloud GPU Image Usage
Introduction
This wiki provides documentation for creating and using a GPU instance in Red Cloud from a Ubuntu (gpu-accelerated-ubuntu-2020-08) or or 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 get users started on using GPU accelerated software with their work quickly with minimal effort.
Start and Setup a Server Instance
- Get started with with the Openstack web interface
- Create an SSH keypair or upload your public key
- Create a custom Security Group if you haven't
Create a Security Group Launch a new Server Instance In the Open Stack Web Interface Dashboard, navigate to Compute > Instances, and click on “Launch Instance”. By Sidebar tab: Details: instance name Source: Boot Source = “Image” Among the list of images, click ⬆️ on the one titled “_____” for Centos 7.8 or “___” for Ubuntu 18.04. Volume Size: increase as necessary. For reference, the Anaconda distribution alone is 19GB. Flavor: Each instance is a Virtual Machine (VM) in the cloud. You can select CPU/RAM/disk configurations (called "flavors") for the VM. We recommend the *t1* for testing the Tesla T4, or the *g1* for testing the Tesla V100 for large runs. If desiring multiple GPUs on one instance, please reach out for special accommodation. Networks: “public” is a good default. Security Groups: Click ⬆️ on the one you made in Step I-C. Click ⬇️ on the “default” group if necessary. Key Pair: Select the key you had uploaded in Step I-B. Click Launch Instance! Note the IP Address of your newly-created Instance. SSH Login into VM In a new terminal window, Chrome extension ssh or Putty, SSH into the VM: ssh -i ~/.ssh/id_rsa/id_rsa centos@<IP_Address_from_previous_step> Assumes that your private key is saved in a file named “id_rsa” at ~/.ssh/id/rsa. CAC Recommends: Monthly security update patching: sudo apt dist-upgrade
What’s Installed on the VM Instance CUDA Check version with yum info cuda (not applicable for Ubuntu) Check CUDA version currently in use with nvcc --version usr/local/cuda is a symlink. If errors, Repoint with ln -s /usr/local/cuda-10.1 /usr/local/cuda as necessary.
CUDA Driver Check version with cat /proc/driver/nvidia/version NVIDIA Drivers Check version with dkms status sudo yum install dkms if “command not found” Check detection of GPU devices by CUDA via NVIDIA’s drivers: nvidia-smi
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, conda info An Anaconda environment should already be activated upon startup, indicated by base preceding each command line prompt.
This “base” environment applies specific settings to your shell for the choice of a Python interpreter and its associated modules and libraries. which python for identifying the environment’s Python location, python -V for versioning Python. 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, docker start `docker ps -q -l` Starts from latest first. To exit container Terminal without exiting or killing the container, CTRL+p then CTRL+q
Testing with Sample Application: with Virtualization only Each of these frameworks can be set up in their own Conda environments. conda env list to see the list of existing usable environments. 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 be ~/anaconda3/envs/tf_gpu/bin/ipython Run sample app within an iPython terminal: ipython
- paste in this sample code:
import tensorflow as tf from tensorflow import keras from tensorflow.keras import layers import numpy as np tf.config.list_physical_devices('GPU')
Output: [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
Run sample app within an iPython terminal: ipython
- paste in this sample code:
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 )))
Expected Output: Tesla T4: Flavor *t1* Tesla V100: Flavor *g1* 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 label 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: https://hub.docker.com/r/tensorflow/tensorflow/ 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: docker ps to check the Container ID (first column) or Name (last column) docker exec -it <Container ID or Name> bash CTRL+p and CTRL+q to exit container without killing it. 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_>
docker exec -it <Container ID or Name from step III-B> 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')
Output: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
Outline Notes begin here ----
How to make a GPU server instance pick the flavor - *t1* for testing V100 for large runs if desiring multiple GPUs on one instance, please reach out for special accommodation reference CAC Wiki existing docs https://www.cac.cornell.edu/wiki/index.php?title=OpenStack#Launch_an_Instance reference _ to be filled in server image ids (assume software libraries installed)
How to login reference CAC Wiki existing docs https://www.cac.cornell.edu/wiki/index.php?title=Red_Cloud#How_to_Create_and_Manage_Red_Cloud_Resources
Run tests that detect and use GPUs, explain what they do, give expected output and basic intro links What is installed and how to list installed software and library versions CUDA yum info NVIDIA Drivers Anaconda conda -V conda # list packages docker images tensorflow in a chosen docker image (Matlab) ... how to detect GPUs nvidia-smi ? python on host (?conda jupyter notebook ssh -L tunnel connection) Docker Keras: both in Docker and at host level App lists devices AND runs on GPU
(Matlab) (Floydhub?)