Dask cuML - Multi-GPU Machine Learning Algorithms

Dask cuML contains parallel machine learning algorithms that can make use of multiple GPUs on a single host. It is able to play nicely with other projects in the Dask ecosystem, as well as other RAPIDS projects, such as Dask cuDF.

Use

As an example, the following Python snippet loads input from a csv file into a Dask cuDF Dataframe and Performs a NearestNeighbors query in parallel, on multiple GPUs:

# Create a Dask CUDA cluster w/ one worker per device
from dask_cuda import LocalCUDACluster
cluster = LocalCUDACluster()

# Read CSV file in parallel across workers
import dask_cudf
df = dask_cudf.read_csv("/path/to/csv")

# Fit a NearestNeighbors model and query it
from dask_cuml.neighbors import NearestNeighbors
nn = NearestNeighbors(n_neighbors = 10)
nn.fit(df)
nn.kneighbors(df)

Dask CUDA Clusters

Using the LocalCUDACluster()

Clusters of Dask workers can be started in several different ways. One of the simplest methods used in non-CUDA Dask clusters is to use LocalCluster. For a CUDA variant of the LocalCluster that works well with Dask cuML, check out the LocalCUDACluster from the dask-cuda project.

Note: It's important to make sure the LocalCUDACluster is instantiated in your code before any CUDA contexts are created (eg. before importing Numba or cudf). Otherwise, it's possible that your workers will all be mapped to the same device.

Using the dask-worker command

If you will be starting your workers using the dask-worker command, Dask cuML requires that each worker has been started with their own unique CUDA_VISIBLE_DEVICES setting.

For example, a user with a workstation containing 2 devices, would want their workers to be started with the following CUDA_VISIBLE_DEVICES settings (one per worker):

CUDA_VISIBLE_DEVICES=0,1 dask-worker --nprocs 1 --nthreads 1 scheduler_host:8786

CUDA_VISIBLE_DEVICES=1,0 dask-worker --nprocs 1 --nthreads 1 scheduler_host:8786

This enables each worker to map the device memory of their local cuDFs to separate devices.

Note: If starting Dask workers using dask-worker, --nprocs 1 must be used.

Supported Algorithms

Nearest Neighbors
Linear Regression

More ML algorithms are being worked on.

Installation

Dask cuML relies on cuML to be installed. Refer to cuML on Github for more information.

Conda

Dask cuML can be installed using the rapidsai conda channel (if you have CUDA 9.2 installed, change the cudatoolkit=10.0 dependency to cudatoolkit=9.2 instead):

conda install -c nvidia -c rapidsai -c conda-forge -c defaults dask-cuml cudatoolkit=10.0

Pip

Dask cuML can also be installed using pip.

pip install dask-cuml

Build/Install from Source

Dask cuML depends on:

dask
dask_cudf
dask_cuda
cuml

Dask cuML can be installed with the following command at the root of the repository:

python setup.py install

Tests can be verified using Pytest:

py.test dask_cuml/test

Contact

Find out more details on the RAPIDS site

Open GPU Data Science

The RAPIDS suite of open source software libraries aim to enable execution of end-to-end data science and analytics pipelines entirely on GPUs. It relies on NVIDIA® CUDA® primitives for low-level compute optimization, but exposing that GPU parallelism and high-bandwidth memory speed through user-friendly Python interfaces.

JohnZed/dask-cuml