Is there any limitation on threads' num for cpu without SMT？

General Usage Performance
1

I’m using 9600KF with 6 cores and 6 threads, the following confuse me:
If I start julia with 6 threads, then the cost of schedule(t::Task) = enq_work(t) increases sharply somtime, while everythings seems ok with threads under 6.

FFTW.fftw_vendor # :mkl
FFTW.set_num_threads(5)
# initialization
a = randn(ComplexF32,512,6)
b = randn(ComplexF32,512,6)
c = similar(a)
# bench function
test(a,b,c) = begin
plan = FFTW.plan_fft(a, 1)
for k = 1:1000
    mul!(b,plan,a)
    tasks = ntuple(1) do i # even i start only 2 tasks
        Threads.@spawn @views c[:,i] .= a[:,i] .* b[:,i]
    end
    @views c[:,6] .= a[:,6] .* b[:,6]
    map(wait,tasks)
end
end
# use Juno.@profiler
@profiler test(a,b,c)

Here’s the profile result:

新建文本文档 ()
新建文本文档 ()1476×695 168 KB

Is this a bug or something else? I found the default threads in Juno for 9600KF is 3, althought it says auto means the number of cpu cores.

2

Well, leaving one CPU core for the operating system is often a good idea…

3

In fact, I only find this if @spawn is used with MKL’s fftw (MKL’s BLAS is ok). :sweat_smile:.
And default FFTW3 is ok…, Maybe I should post a issue to FFTW.jl ? :thinking: