Optimize shared data layout to avoid false sharing

[HadrienG2]

The current implementation has minimal memory overhead, but can exhibit false sharing, which reduces the performance of concurrent reads and writes more than necessary. A more careful data layout could prevent this.

[HadrienG2]

So, what I want here is to align each buffer on a cache line boundary. The back-buffer index and the last buffer index can safely reside on the same cache line, because they are always accessed together.

Two tracks to explore, neither of which is currently available in stable or nightly Rust, are...

• Using repr(align) to manually align structs on a 64-bytes boundary, which is a typical cache line size. Tracking issue for the repr(align) RFC is implement #[repr(align)] (tracking issue for RFC 1358) rust-lang/rust#33626.
• Using a facility dedicated to cache line alignment. This is being discussed as part of Rust issue Add a hardware_destructive_interference_size const to the standard library rust-lang/rfcs#1756.

[mkpankov]

Using repr(align) to manually align structs on a 64-bytes boundary, which is a typical cache line size. Tracking issue for the repr(align) RFC is rust-lang/rust#33626.

It seems it's stable now.

[HadrienG2]

Indeed, and crossbeam-utils will introduce a sane version of CachePadded in their next release. Once they release that, it will probably be the right time to update triple-buffer.

[HadrienG2]

Crossbeam have pushed new releases last month, so I can investigate this now.

[HadrienG2]

Fixed in v3