Peter Andreas Entschev

@mike-wendt @raydouglass based on the discussion from Ops Demo Tuesday, this is a PR we would like to test with Tegra devices when we have it available in CI.

Sounds great, thanks @mike-wendt and let us know if/how we can assist.

> I _think_ the suggestions about RMM pool and UCX are already configured for you if you use `CUDAWorker` class. Is that true? No, neither is currently default. RMM must...

Could you include some more information to help us understand the problem a little better? See below. 1. Could you post the output of `conda list` (if installing the environment...

> * Do the yaml based configuration only get picked up when creating a cluster from a python client _( > from dask_cuda import LocalCUDACluster; cluster = LocalCUDACluster())_ or does...

I'm 100% onboard with this idea @wence- , could you submit a PR to distributed? And thanks for one more neat solution Lawrence-style!

https://github.com/rapidsai/dask-cuda/pull/1021 is now in and should allow specifying `rmm_pool_size=0.95`/`--rmm-pool-size 0.95`, for example. Please let me know if you find any issues with it.

Currently, `DGX` is very limited compared to `dask-cuda-worker`, e.g., it doesn't add the opportunity for device spilling. I guess what you're suggesting is that we make the `DGX` class becomes...

Additionally, we don't want to force users to specify the options that are already available with `dask-cuda-worker`, even though it's technically possible that the user mimics it just with a...

I think having things based on spec-cluster is generally a good idea. Admittedly, I don't know it in details, so I think it would be useful to have some way...