Controlling @threads (or similar) parallelism level

General Usage
1

I have a function which is, abstracting,

Threads.@threads for i in 1:length(chunk)
    memory_intensive_computation(chunk[i])
end

I don’t know a priori how memory intensive the computation will be, though I can assume it’s similar for items within the chunk. Right now, I have to allocate resources with a high enough memory to Julia threads ratio that this rarely runs out of memory, but this strategy is often wasteful. It’d be nice to something more like

memory_per_item = estimate_memory(chunk[1])
n_threads = floor(Int, system_memory / (memory_per_item * safe_inflation_factor))
Threads.@threads n_threads for i in 1:length(chunk)
    memory_intensive_computation(chunk[i])
end

i.e. dynamically choose the right parallelism level. What’s the best way to achieve this kind of pattern?

1 Like
2

(Disregard)

3

Isn’t that just a problem when using Threads.threadid()? Which the code doesn’t use.

1 Like
4

You can do this with OhMyThreads.jl

You’d just write

using OhMyThreads
memory_per_item = estimate_memory(chunk[1])

@tasks for i in eachindex(chunk)
    @set ntasks = floor(Int, system_memory / (memory_per_item * safe_inflation_factor))
    memory_intensive_computation(chunk[i])
end

or

using OhMyThreads
memory_per_item = estimate_memory(chunk[1])

ntasks = floor(Int, system_memory / (memory_per_item * safe_inflation_factor))

tforeach(eachindex(chunk); ntasks) do i
    memory_intensive_computation(chunk[i])
end
3 Likes
5

Indeed. My mistake for attempting a pre-coffee comment :face_with_crossed_out_eyes:

1 Like