How to perform sequential graph partition in distributed training? Or is there a way to manually partition the input graph and store the partition information to files, as dgl.distributed.partition_graph() do? The fact is I have a small input test graph, and I want to assign the nodes to partitions manually. It seems only random or Metis partition can be used currently.
Not supported. Sequentially adding nodes to existing partition involves several modules/parts such as graph.dgl, meta.json which records the node mapping. For now, both random and metis first obtain the node assignments via calls like https://github.com/dmlc/dgl/blob/fa3f2f486b1411f1026e527422943f49288f634b/python/dgl/distributed/partition.py#L883, then generate partitions. You probably could look deep into this part and hack it on your own.
1 Like
I will take a deep look into this. Thanks a lot!
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.