something about loading sort_feature.pt in Paper100MDataset

[zhaoyin214]

it seems strange when loading feature from sort_feature.pt in the Paper100MDataset class in benchmarks/ogbn-papers100M/dist_sampling_ogb_paper100M_quiver.py

i found that the sort_feature has been sorted according to the in-degree order by the statement feature = feature[prev_order] in preprocess.py. in papers100m benchmark, the sorted feature was sorted again. in my opinion, the Paper100MDataset should load feature from feature.pt rather than sort_feature.pt.

is that correct?

[Anlarry]

I also recently learned how quiver accelerates training. I think the purpose of sorting the sorted feature is load balance. Reordering the sorted feature randomly will allow the hottest features to be evenly stored on each GPU.