Julia High Performance - Second Edition

Atomics

We've shown many ways to sum numbers over the course of this book. Let's now try to sum an array of numbers across many threads, hoping to improve performance compared to doing it sequentially on a single thread.

Our first attempt, in the following code, simply tries to run the loop on multiple threads:

function sum_thread_base(x)   r = zero(eltype(x))   @threads for i in eachindex(x)      @inbounds r += x[i]   end   return rend

We can then compare this function against Julia's built-in sum as follows:

julia> a=rand(10_000_000);julia> @btime sum($a)  6.746 ms (0 allocations: 0 bytes)5.000462435093071e6julia> @btime sum_thread_base($a)  1.566 s (5506561 allocations: 84.01 MiB)1.250442324066673e6

The results may be surprising at first glance. Not ...