par_sort_graph: Simplify using Rayon

[progval]

No description provided.

[vigna]

Well, the whole idea was to avoid Rayon as the bridge is horribly slow and prone to contention (that why we wrote ParFoldMap—performance with the bridge was abysmal).

[vigna]

But we can try some benchmarks.

[progval]

But here we're bridging an iterator with only a small number of elements, and the current code spawns a thread for each of these elements.

[vigna]

Yeah, that's the idea—we know the parallelism we want, so we do it. And no Rayon overhead.

[progval]

But Rayon's overhead is negligible on an iterator with less than a thousand items when processing each item takes minutes.

[vigna]

This will be worked out in the current "de-parameterization" of the parallel code.