[polars-rapidsmpf] enable spill to pinned memory

[madsbk]

Add option:

    spill_to_pinned_memory
        Whether RapidsMPF should spill to pinned host memory when available,
        or use regular pageable host memory. Pinned host memory offers higher
        bandwidth and lower transfer performance for device transfers compared to
        regular pageable host memory.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[TomAugspurger]

My only question was whether a boolean will be sufficient, or whether we might need to configure something about the pinned memory resource. But https://github.com/rapidsai/rapidsmpf/blob/331162c5d8eec65352a18596f3d89c2a925db5f0/cpp/include/rapidsmpf/memory/pinned_memory_resource.hpp#L82 mentions that things like a maximum size should be configured outside of the MR, so that shouldn't be needed.

Do we have any opinions on whether this should be the default? Or do we need to collect some benchmark results first?

[madsbk]

My only question was whether a boolean will be sufficient, or whether we might need to configure something about the pinned memory resource. But https://github.com/rapidsai/rapidsmpf/blob/331162c5d8eec65352a18596f3d89c2a925db5f0/cpp/include/rapidsmpf/memory/pinned_memory_resource.hpp#L82 mentions that things like a maximum size should be configured outside of the MR, so that shouldn't be needed.

Yeah, I think a boolean should suffice for now.

Do we have any opinions on whether this should be the default? Or do we need to collect some benchmark results first?

@pentschev did some runs that showed mixed results, which might be because we don’t reuse the BufferResource between evaluate_pipeline() calls. For now, let’s disable pinned memory by default, then see if we can reuse the buffer resource in a follow-up PR. If that consistently improves performance, we can enable it by default.

[madsbk]

/merge