[FEA] Optimize cugraph.uniform_neighbor_sample

[VibhuJawa]

Describe the solution you'd like and any additional context

We should optimize cugraph.uniform_neighbor_sample.

import cudf
import cugraph

df = cudf.read_parquet('/datasets/graph_nn/ogbn_products.parquet')
G = cugraph.Graph(directed=True)
G.from_cudf_edgelist(df,source='src',destination='dst',edge_attr='weight')
batch = G.nodes()[0:4096]

%%timeit 

sampled_g_cugraph = cugraph.uniform_neighbor_sample(G,
                                     start_list=batch,
                                     fanout_vals=[10],
                                     with_replacement=False

332 ms ± 847

Expected Performance : I would expect the performance to be around 8-9x better. More details in internal thread.

[VibhuJawa]

Will be optimized by 50x 332 ms to 25 ms by #2394 .

Of that 25 ms only 6.61 ms seems to spent in the pylibcugraph_uniform_neighbor_sample call so we have more scope there.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[kingmesal]

@seunghwak or @VibhuJawa
Can you confirm if this is complete? Does anything else need to occur to close this?

[seunghwak]

#3269 might be related to this issue no other issue tracking uniform neighbor sampling performance.

[kingmesal]

@VibhuJawa can you confirm whether or not #3269 fixes your issue?

[VibhuJawa]

Lets close this issue. #3269 fixed most things