running DistNeighborLoader causes performance issues

[Harouy]

Hi, i have a question regarding the DistNeighborLoader that I am using alongside OGBN_MAG dataset. I have followed the official documentation and what I have noticed that when using 3 hop neighbourhood like [15,10,5] the performance of the batch geneartion and batch fitting becomes very bad(>80) however limiting the hops to only 2 [15,10] I could clearly see a much better performance about 0.2s fro batch generation and 1s for batch fitting. This issue was not present when working with one machine using only NeighborLoader. However I noticed it appeared when working with multiple VMs. is this a normal issue or could it be linked to my implementation of things(even i followed the original implementation in the repo) with some modifications to be suitable to my environment. for the batch size it is 2000 and for the model it is a simple graphsage model with number of hidden layers (256)

[akihironitta]

what I have noticed that when using 3 hop neighbourhood like [15,10,5] the performance of the batch geneartion and batch fitting becomes very bad(>80) however limiting the hops to only 2 [15,10] I could clearly see a much better performance about 0.2s fro batch generation and 1s for batch fitting.

Hi @Harouy, what is ">80"? Any insights into how changing the number of workers would help with the slowness? Any more relevant information? Also, just for your information, unfortunately, it's not actively maintained due to lack of our bandwidth, and there's even a discussion on whether we should deprecate it or not in #10411.