Machine Learning

[Tutorial: Environment Setup]

I-Ching Tseng
[email protected]
[email protected]
National Taiwan University
March 2022
 Outline
q Overview
q Package Management Tools
q GPU
q Docker
q Conclusion

2
 Overview
q To run a machine learning (ML) model
Ø You have to set up an environment first
Ø Using virtualization or package management tools is a good practice
• You can migrate the code and reproduce the result easily
• Different applications will not affect each other
• If your environment is broken, just create a new environment

q In this tutorial
Ø We will provide some guidelines for setting up environment
Ø We will help you understand the environment
• The software stack
• NVIDIA GPUs

3
 Outline
q Overview
q Package Management Tools
Ø Prerequisites
Ø Conda
Ø Pipenv
Ø Summary
q GPU
q Docker
q Conclusion

4
 Prerequisites
q Package management tools
Ø Help you to manage to environment
Ø Do not manage the GPU driver
q To utilize GPUs, make sure the GPU driver is intalled
Application

Conda/Pipenv

PyTorch

NVIDIA Driver
Software
Hardware
NVIDIA GPU
5
 Conda
q Conda
Ø An open source package and environment management system
Ø Supports Windows, MacOS, and Linux

q We take Anaconda as an example

6
 Quick Start - Anaconda
Steps Linux Command

Install Anaconda with the installer bash Anaconda3-2021.11-Linux-x86_64.sh

(Check the document for details)

Create an environment
(You can replace test_env with conda create -n test_env
your desired environment name)

Install packages conda install -n test_env pytorch torchvision

(You can find the command in torchaudio cudatoolkit=11.3 -c pytorch
the PyTorch official website)

Activate the environment conda activate test_env

Run your application python ml.py

Leave the environment conda deactivate

7
 Pipenv
q Pipenv
Ø A tool that creates and manages a virtualenv

8
 Quick Start - Pipenv
q To know more about Pipenv, please check the document
Steps Linux Command

Install Pipenv with pip3 pip3 install pipenv

pipenv install numpy torchvision torch --index

Install packages https://download.pytorch.org/whl/cu113

Activate the environment pipenv shell

Run your application python ml.py

Leave the environment Ctrl + D

9
 Summary
q To utilize GPU, you must install driver on your host machine

q Using Conda or Pipenv to build environments is recommended

Ø Portable
Ø Reproducible
Ø Applications do not affect each other

q You can stop here if you just want to finish the homework

q Why is PyTorch so convenient?

Ø "We ship with everything in-built (PyTorch binaries include CUDA,
CuDNN, NCCL, MKL, etc.)." [Reference]

10
 Outline
q Overview
q Package Management Tools
q GPU
Ø NVIDIA GPUs
Ø Software Stack
Ø NVIDIA Driver
Ø CUDA
q Docker
q Conclusion

11
 NVIDIA GPUs
q General Purpose Graphics Processing Units (GPGPU)
Ø GPUs are originally designed for computer graphic applications
Ø GPU is good at parallelizing "simple and repetitive" computations
• E.g., matrix multiplication
Ø There are massive matrix multiplication computations in ML models
• We use GPU to accelerate ML model training

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html 12
 Software Stack
Application
Translation Image Classification Regression

Frameworks (Caffe, Tensorflow, PyTorch, etc.)

Generic Convolution Layer cuDNN Optimized Conv. Layer

BLAS Libraries
OpenBLAS MKL2019 cuDNN/cuBLAS

NVIDIA Driver

Hardware
CPU FPGA GPU

13
 NVIDIA Driver
q NVIDIA driver
Ø The software that allows operating systems (OS) to communicate with
GPUs
Ø Includes kernel modules

Frameworks
cuDNN Conv. Layer
cuDNN/cuBLAS
BLAS Lib
cuDNN/cuBLAS CUDA Runtime API

User space
NVIDIA Driver CUDA Driver Kernel space

Hardware
GPU
14
 CUDA
q Compute Unified Device Architecture (CUDA)
Ø "A parallel computing platform and application programming interface
that allows software to use NVIDIA GPUs" [Wikipedia]
q CUDA Runtime API vs. CUDA Driver API
Ø The driver CUDA version must ≥ the runtime CUDA version
Ø Check the driver CUDA version

Ø When we "install CUDA"

• We usually refer to CUDA runtime
• You should check the framework compatibility
• The version should not be greater than the driver CUDA version
• You should choose the runtime CUDA version carefully

15
 Outline
q Overview
q Package Management Tools
q GPU
q Docker
Ø Virtualization
Ø Why using Container?
Ø Contanerization with Docker
Ø Pulling Docker Images
Ø NVIDIA Docker
q Conclusion

16
 Virtualization
q Virtual machine (VM) and container

q You only have to know that

Ø Containers only virtualize software layers above the OS level
• It is a good choice if we only focus on specific hardware (e.g., NVIDIA GPUs)
Ø Containers are relatively lightweight

https://www.docker.com/resources/what-container 17
 Why using Container?
q Containers can virtualize more complex environments
Ø Even if you "only want to train models"
• You may use other frameworks that do not ship with CUDA and cuDNN
• You may need NCCL to perform efficient parallel and distributed training
• You may need to run an old version PyTorch, but the default CUDA version is
too old to communicate with the latest powerful GPU

q Slurm and Kubernetes are popular server management tools

in both academia and industry
Ø Slurm supports singularity container
Ø Kubernetes runs application in Docker containers

18
 Containerization with Docker
q Docker
Ø A platform for you to build and run with containers
Ø Docker installation
• Docker Desktop (for Mac and Windows) runs a VM
q Docker image
Ø A set of instructions for creating a Docker container
q Steps of setting up environment with Docker
Ø Install Docker
• One-time effort
Ø Build/pull an image
• There are lots of built images
Ø Run the container
Ø Run your application

19
 Pulling Docker Images
q Docker Hub
Ø A place for finding and sharing Docker images
• E.g., Docker Hub repository of PyTorch
q Check the Docker Hub and find the image tag
Ø 1.9.1-cuda11.1-cudnn8-devel vs. 1.9.1-cuda11.1-cudnn8-runtime?

Ø Run "docker pull <image_tag>"

20
 NVIDIA Docker (1/2)
q Using GPUs in Docker container makes container less portable
Ø Containers work in user space
• Root privilege only means you can use some privileged system calls
Ø Using NVIDIA GPUs requires kernel modules and user-level libraries
• The CUDA version of the driver user-space modules must be exactly the same
as the CUDA version of the driver kernel modules
• The runtime CUDA version can be smaller than the driver CUDA version
Ø The host driver must exactly match the version of the driver installed in
the container

q We should use NVIDIA Docker

Ø Install NVIDIA Docker
Ø You do not have to install the NVIDIA driver in the container

https://github.com/NVIDIA/nvidia-docker 21
 NVIDIA Docker (2/2)
q Steps
Ø Install the latest NVIDIA driver
• One-time effort
Ø Install NVIDIA Docker
• One-time effort
Ø Build/pull an image
Ø Run the container
Ø Run your application

22
 Outline
q Overview
q Package Management Tools
q GPU
q Docker
q Conclusion

23
 Conclusion
q Whether or not you virtualize your environment
Ø You must install the NVIDIA driver on the host to utilize NVIDIA GPUs
Ø The runtime CUDA version must be less than or equal to the driver
CUDA version

q If you want to use NVIDIA GPUs in containers

Ø Using NVIDIA Docker makes your life easier
• You do not need to install NVIDIA drivers in containers
• Containers are more portable
Ø You only have to pull the built Docker image from Docker Hub
• You do not have to set up CUDA, cuDNN, and frameworks yourself
• This is useful especially when the environment is complex

24
 Q&A

Thank You!

25