Spark Rapids ML enables GPU accelerated distributed machine learning on Apache Spark. It provides several PySpark ML compatible algorithms powered by the RAPIDS cuML library, along with a compatible Scala API for the PCA algorithm.

These APIs seek to minimize any code changes to end user Spark code. After your environment is configured to support GPUs (with drivers, CUDA toolkit, and RAPIDS dependencies), you should be able to just change an import statement or class name to take advantage of GPU acceleration.

Python

# from pyspark.ml.feature import PCA
from spark_rapids_ml.feature import PCA

pca = (
    PCA()
    .setK(3)
    .setInputCol("features")
    .setOutputCol("pca_features")
)
pca.fit(df)

Scala

// val pca = new org.apache.spark.ml.feature.PCA()
val pca = new com.nvidia.spark.ml.feature.PCA()
  .setK(3)
  .setInputCol("features")
  .setOutputCol("pca_features")
  .fit(df)

The following table shows the currently supported algorithms. The goal is to expand this over time with support from the underlying RAPIDS cuML libraries. If you would like support for a specific algorithm, please file a git issue to help us prioritize.

Note: Spark does not provide a k-Nearest Neighbors (k-NN) implementation, but it does have an LSH-based Approximate Nearest Neighbor implementation. As an alternative to PCA, we also provide a Spark API for GPU accelerated Uniform Manifold Approximation and Projection (UMAP), a non-linear dimensionality reduction algorithm in the RAPIDS cuML library. As an alternative to KMeans, we also provide a Spark API for GPU accelerated Density-Based Spatial Clustering of Applications with Noise (DBSCAN), a density based clustering algorithm in the RAPIDS cuML library.

• For PySpark (Python) users, see this guide.
• For Spark (Scala) users, see this guide.

GPU acceleration can provide significant performance and cost benefits. Benchmarking instructions and results can be found here.

We welcome community contributions! Please refer to CONTRIBUTING.md to get started.