RAPIDS

What is RAPIDS?

RAPIDS is a suite of open source software libraries developed by NVIDIA to accelerate data science pipeline on GPUs. Each component is modeled after its CPU counterpart with minimal code change for the user.

Components

Component CPU API Note
cupy NumPy, SciPy Technically not part of RAPIDS
cuDF Pandas
cuML Scikit-learn
cuGraph NetworkX
cuSignal SciPy Signal
cuSpatial GeoPandas
cuxfilter crossfilter

We will focus on the first three components in this introductory workshop. The others are more specialized.

Prerequisites

  • NVIDIA GPU with compute capability 6.0+ (Pascal)
  • CUDA 11

Installation

The rapidsai module is available on Rivanna. It is backed by a container that we built. (The container image is 40% smaller than the official one. Learn more about containers to understand why.)

To install on your own machine please visit https://rapids.ai/start.html.

Usage

JupyterLab

Select the “RAPIDS x.y” kernel.

Command line

module load singularity rapidsai

Exercises

A clone of https://github.com/rapidsai/notebooks is located at /project/apps_data/rapids/notebooks. The full size is near 1 GB. For this workshop, just copy the cudf and cuml subdirectories:

cp -r /project/apps_data/rapids/notebooks/repos/{cudf,cuml} ~

We have also prepared a notebook cudf.ipynb under /project/apps_data/rapids.

  1. In cudf.ipynb, compare the performance of pandas and cudf starting from N = 100. Explore the behavior by varying N. Beyond which order of magnitude does cudf outperform pandas?

  2. Which method’s relative performance remains fairly constant? In other words, the ratio of the pandas execution time to the cudf execution time does not change much with respect to N.

  3. Which method has the highest performance boost using cudf?

  4. Repeat all of the above for other data types.

  5. In cuml/notebook/kmeans_demo.ipynb, compare the performance of scikit-learn and cuml by varying N. Beyond which order of magnitude does cuml outperform scikit-learn?

  6. There is a cell that checks the accuracy of cuml versus scikit-learn. Is this necessary?

  7. Feel free to explore other notebooks.

  8. General question: Why does the performance of RAPIDS depend on N? Why does the CPU API outperform RAPIDS when N is not big enough?

Remark: JupyterLab vs batch job

The JupyterLab environment is interactive which is great for debugging and testing. However, if the queue is busy you may need to wait for a long time. If your code can be executed non-interactively, we recommend converting it into a Python script so that you can submit it as a batch job.

Converting a notebook into a Python script

The following command will convert your notebook mynotebook.ipynb into mynotebook.py.

module load anaconda
jupyter nbconvert --to script mynotebook.ipynb

You may need to comment out Jupyter magic commands (e.g. %%time) before converting.

Batch job

Prepare a SLURM script job.slurm:

#!/bin/bash
#SBATCH -A mygroup             # your allocation account
#SBATCH -p gpu                 # partition
#SBATCH --gres=gpu:1           # number of GPUs
#SBATCH -N 1                   # number of nodes
#SBATCH -c 1                   # number of cores
#SBATCH -t 10:00:00            # time

module purge
module load singularity rapidsai

# change x.y to the actual version
singularity run --nv $CONTAINERDIR/rapidsai-x.y.sif mynotebook.py

Submit the job via sbatch job.slurm.


References and further reading