[REVIEW] MG (Batch) Betweenness Centrality

[xcadet]

This work adds Multi-GPU Batch Betweenness Centrality for both Vertices and Edges:

• Data Replication is operated once for the graph using dask and RAFT.
• Task Distribution is operated at the Cython level

This version allows to use the same Graph for both Single GPU algorithms and algorithms supporting Multi GPUs Batch
processing.
Graphs can be loaded the same way they are with cudf for single GPU, and can leverage Multi GPUs Batch Processing by using G.enable_mg_batch() where G is a Graph or DiGraph. By enabling mg_batch edge list and / or adjacency lists are replicated across device using dask and RAFT's communicator. When enabled and when adding representations to the graph, replications of the representations are operated. Replication of the transposed adjacency list is not yet implemented.

This also adds functions for replciation of cudf.DataFrame and cudf.Series over multiple GPUs, and a notebook to show how to use Multi GPUs Batch Processing for Betweenness Centrality

A boolean parameter is added at the C++ level to BFS to disable MG BFS path to be taken when called from MG Batch BC.

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[afender]

Looks good. Added ome minor questions/comments.

Please add a note on Batched processing in the .rst file under doc like after the Consolidation section.

[xcadet]

rerun tests

[afender]

Looks good. Need to investigate graph replication time compared to peak BW expected perf

[seunghwak]

Just add a FIXME comment, and thanks for the contribution!

[BradReesWork]

rerun tests

[xcadet]

rerun tests

[BradReesWork]

rerun tests