The allocations introduced by Threads.@threads were also one of the reasons why I started an older discussion on the overhead of this macro. Finally, we switched to @batch from Polyester.jl (which wasn’t available back then) in Trixi.jl, our package for hyperbolic PDEs. The macro @batch includes a check whether only a single thread is used; if so, it will run the plain serial code, allowing you to see allocations in your inner functions. If multiple threads are used, which can be really helpful when developing code. If multiple threads are used, it usually creates much less allocations than Threads.@threads.
However, let me warn you that @batch isn’t as general as Threads.@threads. Additionally, we ran into some interesting bugs, but Chris Elrod was really helpful and fixed all the problems very fast (as always!).
3 Likes