Dask-cuda: Add minimal example of dask-cuda command line cluster setup

[beckernick]

In the docs, we have a clear example of setting up a LocalCUDACluster. We also list several additional arguments to the "Command Line Tool", which corresponds to the Dask CLI. However, it's not obvious from the dask-cuda documentation what needs to change relative to standard usage of the Dask CLI. We have an example of using dask-cuda-worker in the UCX integration setup, but it's quite possible a user wouldn't look at the UCX integration if they aren't already planning to use UCX.

It would be valuable if we included a minimal example of setting up a multi-node cluster (or equivalent single-node setup) with dask-cuda-worker from the command line, like in the Dask docs. This could live in either dask-cuda or dask. This documentation would be a valuable asset to share with users interested in testing multi-node or from the command line.

[pentschev]

Just recently @charlesbluca added some code to try and address specifically that, which now lives in https://github.com/rapidsai/dask-cuda/tree/branch-0.19/examples/ucx . Does that cover your question, if not, what else could we do to improve it?

[charlesbluca]

Maybe it would be better to duplicate some of the general usage info in the UCX section in the Worker section, so that it's more visible? We can always slim down the UCX section later on if we find that parts of it are redundant.

I'm also happy to expand upon the standalone utilities - if we made the use of UCX an option, then they could function as a minimal example of a standard multi/single-node set up.

[beckernick]

@charlesbluca I think that approach would be a great idea

[charlesbluca]

Currently drafting some changes to address the docs aspect of this in #561 - in general, I am leaning towards putting basic usage examples in the first few pages, and moving more advanced configuration to standalone examples in the docs.

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[jakirkham]

Is this resolved with PR ( https://github.com/rapidsai/dask-cuda/pull/561 ) or was there more work needed here?

[charlesbluca]

I think that's sufficient - we could always potentially add an example script like those made for UCX.

@beckernick did you get a chance to check out the changes to the docs?