Scoping in loops and multi-threading

New to Julia
1

Hello everyone,
I just started using multi-threading, and have a basic question about multi-threading with loops. This might be a bit of a duplicate on some other questions here (like this: PSA: Reasoning about scope rules and multithreading) but I was still a bit confused. It would be great to get some advice on this!

Here is a rough outline of the code I’m running:

@threads for i = 1:N
data = full_data[agent_id .== i,:] 
# thus subset of data used varies by iterations
iter = 1.0
tol = 1e6

   while ( (iter < 5000) && (tol > 1e-8) )
    #-----some code here using data subset --------#
    #-----create new_vec------#

    tol = norm(new_vec - old_vec)
    iter = iter + 1
   end

end

You can think that within each iteration of the for-loop, I’m running a while loop. In each iteration of the while loop, I’m creating a vector called new_vec, and comparing that with the previous while-loop’s created vector (old_vec). If the difference is small enough, I stop the while-loop, and move on to the next iteration of for-loop. Note that I am not writing to data or full_data, I just use them to create new_vec.

I’m multi-threading the for-loop. Each iteration of the for-loop does not depend on other iterations, i.e. new_vec for i = 100 doesn’t influence new_vec for i = 500.

My worry is that if I do this, then there could be some sort of data race being created for iter and tol. That is, the value of iter in one i could end up in the value of iter in another i, and similarly for tol. Is this an issue? Or are the scopes here somehow for-loop iteration specific, i.e. tol and iter in each iteration of for-loop is untouched by others?

I also did the following but I’m not sure if it would help.

@threads for i = 1:N
data = full_data[agent_id .== i,:] 
# thus subset of data used varies by iterations
local iter = 1.0
local tol = 1e6

   while ( (iter < 5000) && (tol > 1e-8) )
    #-----some code here using data subset --------#
    #-----create new_vec------#

    tol = norm(new_vec - old_vec)
    iter = iter + 1
   end

end

Does putting in the local at the beginning limit the scope of tol and iter to that specific for-loop iteration?
Any suggestions on how to improve the code would be appreciated too. Thanks!

2

that is true even without multithreading

their scope is already local to each iteration of the loop. You probably want to add a local old_vec, beause the same is valid to the iterations of the while loop, and if you do not add local old_vec (or initialize it with a value) outside that while loop, its value won’t be preserved from one iteration to the other.

2 Likes