Hi,
I am trying to use multi-threading in a monotonic dynamic way (in the sense of openmp). By monotonic , I mean that if a thread exected iteration i then the thread must execute iterations larger than i subsequently.
So for example, 2 threads and, if we have i = 1, 2 , 3 … n, then thread 1 will execute first i =1 and thread 2 will execute i = 2. Whatever thread finishs first it will execute i=3 and so on…
1/ is it how Threads.@threads work?
2/ if not then how do I do that for n threads in Julia?
Thank you
Kind regards
I was wondering whether there is maybe a more appropriate place to ask these questions?
Thank you
No. This example illustrates it:
using .Threads
lk = ReentrantLock()
disorder = Int[]
@threads for i in 1:20
@lock lk push!(disorder,i)
end
The content of array disorder is usually not monotonic.
Here’s one way it can be achieved:
iter = Atomic{Int}(1)
order = Int[]
@threads for _ in 1:20
i = atomic_add!(iter,1)
@lock lk push!(order, i)
end
(Not that the array order is guaranteed to be monotonic, some thread can be delayed, though often not in this simple example).
1 Like
Thank you very much. I need to study your code because there is a lot of things I am not familiar with here.
Am I right in saying that yhe value in order should be ordered from 1 to 20? I am using Julia 1.8.1 on windows and sometimes it is indeed ordered but sometimes not…especially on a fresh session.
Also if I increase the loop iteration from 20 to let’s say 600. It actually occurs more frequently.
No. The iterations (i.e. the is) are started in order, but not necessarily finished in order. (The lock-wait isn’t first-in-first-out). Within each thread there will be monotonicity, however.
Strange, I am probably doing something wrong then. I don’t see performance improvement I see on the C side. I will post some code eventually. For now, I will accept your answer. Thank you for your help.