How to use all of the cpu cores in IJulia using broadcasting function?

General Usage Performance
1

I am relatively new to Julia and I have been exploring how to parallel process functions in IJulia. I have created a Jupyter notebook for 128 cores using the directions given here. I can confirm that it has imported all of the requested cores using the .nthreads. Now I have a function that calls a bunch of functions inside ( My actual function is very long, so for instance for an explanation as shown).

function calculate(variables)  #variables is a vector
   x1,x2,x3,x4,x5,x6 = variables
   value1 = funct1(x1,x2,x3) # this function returns a real value
   value2 = funct2(x3,x4,x3)  # this function returns a real value
   value3 = funct3(x3,x5,x6)  # this function returns a real value
   return funct4(value1, value2, value3)  # this function returns a vector.
end

numbers = [[rand() for i=1:6] for j=1:100000000]
answer =  calculate.(numbers) # I get the expected answer but CPU is not utilised well.

As I execute these codes and when I check the “system monitor”, I see that only one core is being used at a time. Please help how I can use all of the cores. I hope I make some sense. I have also tried FLOOP and multithreading packages but I see that all cores to 100% and never give me an answer. Thanks for your help.

P.S. I use Ubuntu 22.

1 Like
2

Where are you actually multithreading your code? Why do you think this is an IJulia issue, do you see more cores being used when running the code in the REPL?

1 Like
3

calculate.(numbers)

does not utilize parallelism and its functional version is

map(calculate, numbers)

and ThreadsX package includes the multi-thread version of this function as

ThreadsX.map(calculate, numbers).

Another way is to use Threads package and @threads but in this case you implement a vector of answers like

answers = ...
Threads.@threads for i in 1:n
    answers[i] = ...
end

in an imperative way.

3 Likes
4

Sorry, looks like my code is not multithreading. I am a bit confused I guess. The confusion between explicit for loop and broadcasting or map function. I am still not able to identify the merits/demerits of the latter. Thanks.

5

Not sure what you mean but you should take a look at the multi-threading documentation.. The idea is that you have to 1) launch Julia with the number of threads you want, and 2) decorate your code with either Julia’s low level @threads or use a library like ThreadsX.

You don’t get parallelism for free just because you use broadcasting.

2 Likes