Tiny experiment to check the effect of cache-line sharing for concurrent libraries. Note: Only works on MacOS right now. Working on making this portable.
The library has one useful function -
val make : int -> int Atomic.t
This creates an atomic integer which is aligned to a cache line boundary and padded such that it does not share the cache line with any other element.
The benchmark has an SPSC queue taken from - https://github.com/bartoszmodelski/ebsl
The benchmark consists of spawning 2 domains where one domain pushes N elements to the queue and another pops N elements.
dune build
/usr/bin/time -v _build/default/bench/main.exe 0
/usr/bin/time -v _build/default/bench/main.exe 1
which runs cases where unaligned and aligned integers are used respectively.
On my machine the results are -
* 7.28s (unaligned)
* 5.62s (aligned)
There is a reduction of 22.80% in the wall-clock time for the aligned case over the unaligned case.
htop was manually inspected to observe about 192% utilization for the bench
process. User time reported by time is rougly 2x of wall-clock time, thus
double checking that 2 domains were concurrently resident throughout the process.