Dataframe.compute hangs during iterating a dataset

Hi, I’m new to dask suite, and I met this tricky problem when I used NVTabular, which is based on dask, to build a dataloader for my pytorch model.
I posted a discussion in dask’s repo. Could anyone help me fix this problem?

Hi, I discover some new information about the question by inspecting the dashboard.

I want to post it there, but this forum only allows me to upload only one image. So I append the screenshot in the discussion.

Is there any reason during executing dask.dataframe.compute() could trigger the CPU utilization fall?