Training GNN takes too much CPU

Questions
#1

I found that while training GNN on CUDA using mini-batch, CPU usage is very high (~1900%).
How can I fix this?
image

#2

Usually this is because it uses OpenMP in places other than model propagation (e.g. neighbor sampling is a common case).

What is your use case? If you were indeed training a node classification model with NodeDataLoader, it supports GPU-based neighbor sampling by placing in the graph on GPU.

#3

High utilization is by design. Why do you prefer lower utilization?

#4

Thanks for your reply.
My DGL version is 0.5.3, neighbor sampling requires the graph to be on CPU.

#5

Thanks for your reply.
I thought this was caused by some bugs in my code.

#6

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.