Ian Henriksen

After thinking about this a bit more, I think the only way to do this is to allow the end-user to pass a list of names that they want excluded...

Partially fixed in https://github.com/ut-parla/Parla.py/pull/41.

On the other hand not detecting duplicates provides an easy way for a user to say "I'm also going to allocate something like this other thing."

Yah, the idea is to allow this, not require it. It's one of those "you can tune this if you need to" kind of stories. I haven't given a lot...

If someone wants to take the lead on this, it may be a good way to get familiar with the internals of the current scheduler.

This is likely really easy to implement. It's just a matter of making it happen. Not before the deadline though.

See also: https://github.com/numba/numba/issues/5137

Actually that's only semi-related. That's about using the per-thread default streams introduced in CUDA 7 in their CUDA module. It's not suggesting interop between numba kernel calls and cupy stream...

https://github.com/numba/numba/issues/4797 is semi-related. I'll try creating an issue upstream to see what they say though.

Submitted upstream as https://github.com/numba/numba/issues/6921. We'll see if they can give us some guidance.