Multi-GPU with Dask-cuDF#

cuDF is a single-GPU library. For Multi-GPU cuDF solutions we use Dask and the dask-cudf package, which is able to scale cuDF across multiple GPUs on a single machine, or multiple GPUs across many machines in a cluster.

Dask DataFrame was originally designed to scale Pandas, orchestrating many Pandas DataFrames spread across many CPUs into a cohesive parallel DataFrame. Because cuDF currently implements only a subset of Pandas’s API, not all Dask DataFrame operations work with cuDF.

The following is tested and expected to work:

What works#

  • Data ingestion

    • dask_cudf.read_csv

    • Use standard Dask ingestion with Pandas, then convert to cuDF (For Parquet and other formats this is often decently fast)

  • Linear operations

    • Element-wise operations: df.x + df.y, df ** 2

    • Assignment: df['z'] = df.x + df.y

    • Row-wise selections: df[df.x > 0]

    • Loc: df.loc['2001-01-01': '2005-02-02']

    • Date time/string accessors: df.timestamp.dt.dayofweek

    • … and most similar operations in this category that are already implemented in cuDF

  • Reductions

    • Like sum, mean, max, count, and so on on Series objects

    • Support for reductions on full dataframes

    • std

    • Custom reductions with dask.dataframe.reduction

  • Groupby aggregations

    • On single columns: df.groupby('x').y.max()

    • With custom aggregations:

    • groupby standard deviation

    • grouping on multiple columns

    • groupby agg for multiple outputs

  • Joins:

    • On full unsorted columns: left.merge(right, on='id') (expensive)

    • On sorted indexes: left.merge(right, left_index=True, right_index=True) (fast)

    • On large and small dataframes: left.merge(cudf_df, on='id') (fast)

  • Rolling operations

  • Converting to and from other forms

    • Dask + Pandas to Dask + cuDF df.map_partitions(cudf.from_pandas)

    • Dask + cuDF to Dask + Pandas df.map_partitions(lambda df: df.to_pandas())

    • cuDF to Dask + cuDF: dask.dataframe.from_pandas(df, npartitions=20)

    • Dask + cuDF to cuDF: df.compute()

Additionally all generic Dask operations, like compute, persist, visualize and so on work regardless.

Developing the API#

Above we mention the following:

and most similar operations in this category that are already implemented in cuDF

This is because it is difficult to create a comprehensive list of operations in the cuDF and Pandas libraries. The API is large enough to be difficult to track effectively. For any operation that operates row-wise like fillna or query things will likely, but not certainly work. If operations don’t work it is often due to a slight inconsistency between Pandas and cuDF that is generally easy to fix. We encourage users to look at the cuDF issue tracker to see if their issue has already been reported and, if not, raise a new issue.