What is julia doing with your threads?

[carstenbauer]

Somewhat related discussion around that the main thread is in the :default or :interactive pool depending on if one starts Julia with interactive threads.

[mbauman]

@PetrKryslUCSD

s = Float64[]
Threads.@spawn let r = $ch[1], i = $ch[2]
    push!(s, work(r))
end

That’s not thread-safe. Breaking thread-safety not only affects accuracy, but it often ends up harming performance, too. I highly suspect that (or GC, as in the other thread that reminded me of this one) as being the problem here.

[PetrKryslUCSD]

Somehow I don’t see how the compiler could mind the access to s.
Otherwise it would know it was not thread-safe and it could be made
be thread-safe by the compiler… But I don’t think it does.

[mbauman]

The problem isn’t the access, it’s the mutation. The threads will trample over eachother as they all try to grow the memory.

This is definitely a problem:

julia> function f(s, n)
           Threads.@threads for i in 1:n
               push!(s, i)
           end
           return s
       end
f (generic function with 1 method)

julia> f(Float64[], 10000)
4630-element Vector{Float64}:
    1.0
    2.0
    3.0
    ⋮
 6249.0
 6250.0

julia> length(f(Float64[], 10000))
3162

julia> length(f(Float64[], 10000))
6250

julia> length(f(Float64[], 10000))
1678

julia> minimum(f(Float64[], 10000))
5.2150196471021e-310