Inconsistent CPU utilisation in @threads loops

[carstenbauer]

@ysh

I gave OhMyThreads.jl a try. Actually, I would like to pre-allocate one buffer matrix per thread instead of allocating as many as there are iterations. So I did this:

using OhMyThreads: TaskLocalValues.TaskLocalValue, tforeach, StaticScheduler
BLAS.set_num_threads(1)
Random.seed!(1)
function testp(M, n)
    ws_type = typeof(EigenWs(M, rvecs=true))
    ws = TaskLocalValue{ws_type}(() -> EigenWs(M, rvecs=true))
    ms = TaskLocalValue{typeof(M)}(() -> similar(M))
    tforeach(1:n, scheduler=StaticScheduler()) do i
        m = ms[]
        m .= M # we're doing a simple copy, but m could depend on i
        LAPACK.geevx!(ws[], 'N', 'N', 'V', 'N', m)
    end
end

Why do you make ms a TaskLocalValue and then perform a (strange) copy in the “parallel loop”? In your OP, M was just a single input matrix?

(I think you might be confusing input data with temporary task-local buffers.)

@ysh

I suppose this is because ms[] or ws[] can trigger GC.

Not really.

@ysh

Is there a way to make each thread access an element of a pre-allocated buffer? Something like ms[threadid()], but the docs say threadid() is not safe to use.

If possible, you shouldn’t think about threads but tasks as Julia implements task-based multithreading. I recommend you take a look at JuliaUCL24/notebooks/Day3/2_multithreading.ipynb at main · carstenbauer/JuliaUCL24 · GitHub. Instead of accessing a pre-allocated element per thread you should likely be accessing a pre-allocated element per task and that’s what TaskLocalValue gives you.

But to answer your question, the following creates nthreads() tasks each running on a different thread and accessing an element of a pre-allocated buffer without using threadid().

data = rand(nthreads())
tforeach(1:nthreads(); scheduler=StaticScheduler()) do tid
    println(data[tid])
end

Don’t think that’s what you want to do though.