Inconsistent CPU utilisation in @threads loops

General Usage Performance
, , ,
18

Apart from the obvious (one is a loop macro the other one isn’t), tforeach gives you

  • load balancing by default (it creates 2*nthreads() tasks by default)
  • more control over the “chunking”, that is, how the interval is divided into tasks (@threads :dynamic/:static always creates nthreads() tasks)

We want to add more features, like new schedulers, perhaps automatic pinning of threads, and more in the future. As for the “macro API”, it gives us some flexibility to make some patterns, like using TLS, a bit simpler (although we generally prefer the higher order function API).

OhMyThreads.jl also supports reductions, which @threads doesn’t (although @distributed does for multiprocessing). The @threaded macro will automatically expand to tmapreduce (rather than tforeach) if you provide a reducer argument, e.g.

julia> @threaded reducer=(+) for i in 1:3
           sin(i)
       end
1.8918884196934453

But now it’s time for me to stop (1) promoting our package and (2) derailing this thread any further. :slight_smile:

3 Likes