[FEA] Add a pinned memory resource

[jrhemstad]

Is your feature request related to a problem? Please describe.

There's currently no memory resource for allocating pinned memory (e.g., cudaHostAlloc).

Describe the solution you'd like

There should be a pinned_memory_resource.

Additional context
Inspired by rapidsai/cudf#2872 (comment)

[UmashankarTriforce]

I would like to work on this issue. May I know more details on this issue?

[jrhemstad]

You can look at the cuda_device_resource as an example.

The pinned_memory_resource should be similar to this, but use cudaHostAlloc and cudaFreeHost instead of cudaMalloc/cudaFree.

Then update the tests to test the new resource: https://github.com/rapidsai/rmm/blob/branch-0.10/tests/mr/mr_tests.cpp#L88

[UmashankarTriforce]

Alright thanks for the information! I'll work on it and send a PR 😄

[jakirkham]

This was done in PR ( #272 ), right?

[jrhemstad]

This was done in PR ( #272 ), right?

The difference is in a host_memory_resource vs. a device_memory_resource. Pinned memory (like managed) is kind of an oddball in that it can be host or device memory.

[jakirkham]

Oh, I didn't realize pinned memory could be treated like device memory. Had assumed it was just page locked host memory. Is that not right? What am I missing here?

[harrism]

It is just that, but CUDA devices are able to access it directly as if it were device memory.
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#page-locked-host-memory

[jakirkham]

Would it make sense to use pool_memory_resource with a pinned_memory_resource then? Or am I getting carried away? 😅

[harrism]

Conceptually yes. However as Jake pointed out in another issue/PR, the problem is we have specific device_memory_resource tests which check that the returned pointer is a device pointer. I think we could actually expand that to instead just check that the memory is device accessible. To do that we would just check that the type is NOT cudaMemoryTypeUnregistered, and then device MRs could be used with device memory, managed memory, or registered (pinned) host memory.

[github-actions]

This issue has been marked rotten due to no recent activity in the past 90d. 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.

[github-actions]

This issue has been marked stale due to no recent activity in the past 30d. 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 marked rotten if there is no activity in the next 60d.

[harrism]

Superseded by #618