RSN 38 - Support for 24.02 cugraph Users

Overview

RAPIDS 24.02 cuGraph users of graph degree functions on undirected multi-GPU distributed graphs will need to apply a temporary workaround to compute correct results.

Example

...
# create a distributed undirected graph from a dask DataFrame
dG = cugraph.Graph(directed=False)
dG.from_dask_cudf_edgelist(ddf, source="src", destination="dst")
...
result_df = dG.degree()
# Temporary workaround for 24.02 only:
#   MG degree() results for undirected graphs are doubled so divide by 2
result_df["degree"] //= 2

Impact

Impact is limited to RAPIDS 24.02 cugraph users using the graph degree functions (G.degree(), G.in_degree(), G.out_degree()) on undirected, multi-GPU distributed graphs.

Background

A bug was noticed late in the 24.02 release process for cugraph users using the degree functions (G.degree(), G.in_degree(), G.out_degree()) on undirected multi-GPU distributed graphs. The bug causes the returned values to be doubled, and users of these functions for these graph types must apply a workaround to divide the result by two in order to obtain the correct values as seen in the example above.

This workaround should only be used for RAPIDS releases 24.02, and only for undirected multi-GPU distributed graphs.