Consider switching to the deque from crossbeam about rayon HOT 3 CLOSED

The strategy which the deque crate uses is to store old buffers in a linked list, doubling in size whenever growth is needed. This ends up converging at 2N space usage for N elements, so it's not an unbounded memory leak. You could alternatively adopt a strategy (requiring modification of the deque crate) to free the extra buffers when the deque provably has no jobs available. A pop operation returning "None" might even be enough proof.

from rayon.

Note: the other, probably more important reason that crossbeam's deque may ultimately be worth trying out is that it's soon going to go from all SeqCst operations to weaker operations in many cases, which will result in far fewer memory barriers. I'm very curious to see how much a difference that winds up making in practice for a consumer like Rayon. (We're still working out some kinks, though.)

from rayon.

Seems like we won't do this (see conversation on #195).

from rayon.