[FEA] Ability to train larger datasets from managed memory (atleast for RF)

[teju85]

Is your feature request related to a problem? Please describe.
If the dataset is large enough not to fit in the GPU memory, we should be able to provide an option for the users to train RF model by keeping the dataset entirely on managed-memory.

This can be a more generic request for all algos in cuML. However, to begin with, we could limit ourselves to RF algo.

Additional context
We should certainly study performance implications of this and document those, if any observed.

[dantegd]

I believe @cjnolet has done some performance analysis of Naive Bayes with managed memory (correct me if I'm misremembering).

But it would be excellent to have a comprehensive list of results for different types of algorithms! Another thing that will be very useful that is coming very soon is using FAISS 1.7 (#3509 ) that allows for plugging a memory manager, so we should be able to have a single pool e2e for algorithms that use FAISS as well.

[cjnolet]

Here's the original issue about pluggable management filed with the FAISS team: facebookresearch/faiss#1203

[github-actions]

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.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 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.