Remove the pageable pool

[revans2]

This is the first step in trying to get host memory under control. I ran performance tests and didn't see any difference in the median performance of NDA.

I didn't remove the config for the pageable pool size yet, but I am debating if I should. The original plan was to not remove it, but keep it for backwards compatibility. I thought it could be used outside of spill, but it cannot, so I think I want to adjust things with it.

[revans2]

build

[revans2]

I adjusted the copyrights and ran NDS at scale factor 200 on my local desktop, but really configured to spill to host memory (8 GiB of GPU Memory, 32 GIB of Spill, 2 GiB of pinned, and 6x parallelism). I still need to run one more set of benchmarks to see what happens with a larger pageable pool size, but with just a 1 GiB default pool there was a 2% performance regression. I personally think that this is acceptable for now. But I will update the host memory limits to have us look at adding a pageable store back in when we are done.

[revans2]

build

[revans2]

I ran tests with pinned memory instead of pageable memory, and it the pinned memory ended up being 10% faster than the pageable memory for the same run, with spilling. The advantage of pageable is that the pinned memory must be materialized, but pageable does not. I don't think I will file a follow on issue for pageable unless we see a specific need for something like it.