Modify MNMG louvain to support an empty vertex partition #1744
Conversation
Codecov Report
@@ Coverage Diff @@
## branch-21.10 #1744 +/- ##
===============================================
Coverage ? 59.75%
===============================================
Files ? 77
Lines ? 3523
Branches ? 0
===============================================
Hits ? 2105
Misses ? 1418
Partials ? 0 Continue to review full report at Codecov.
|
BradReesWork
added
3 - Ready for Review
improvement
| vertex_t* clustering) | ||
| { | ||
| if (graph_view.get_number_of_local_vertices() > 0) | ||
| CUGRAPH_EXPECTS(clustering != nullptr, "Invalid input argument: clustering is null"); |
There was a problem hiding this comment.
Is this really an invalid use? I assume it is valid (even though hugely inefficient) to ask cuGraph to compute Louvain for a tiny graph using many GPUs (e.g. # GPUs > # vertices).
There was a problem hiding this comment.
If the number of local vertices is 0 then the clustering pointer is never referenced, nothing is required to be written.
The issue that triggered this is that if the graph is poorly distributed (karate over 8 GPUs apparently ends up with at least 1 cluster which has no vertices assigned to it) then the python code creates a cudf column with 0 entries. This results in one of the calls using clustering == nullptr, which was failing. This check will allow clustering to be nullptr if there are no vertices assigned to this GPU. If there are vertices assigned to this GPU it will fail if clustering is nullptr
There was a problem hiding this comment.
Ah... I see. I misunderstood the code and thanks for the explanation.
|
@gpucibot merge |
Discovered in MNMG testing.
Closes #1713