GPU accelerated graphs

[johnlees]

Could look at using https://github.com/rapidsai/cugraph

[johnlees]

Triangles: cugraph.community.triangle_count.triangles(G)
Denominator: degree = G.degree(); sum([d * (d - 1) for d in degree)
Transitivity = triangles/denominator
(see https://networkx.org/documentation/stable/_modules/networkx/algorithms/cluster.html#transitivity)

Cliques: https://docs.rapids.ai/api/cugraph/stable/api.html?highlight=degree#module-cugraph.community.ktruss_subgraph

[nickjcroucher]

Should this be added under a --gpu-network flag? Could try adding these functions one-by-one that way I suppose, once the big PR is merged.

[johnlees]

It's supposed to be completely interoperable with networkx (perhaps unfortunate now we've completely removed that), but I think a third GPU flag to enable this would be very sensible

[nickjcroucher]

So networkx-type code needs to be added back in? At least we can retrieve the code.

[johnlees]

Yeah. I think we'll keep it as graph tools, and just have functions which convert to/from cugraph, and then the time-consuming ones (cliques, network score) should be fairly small drop-ins