Is this normal things about MvNormal?

General Usage
,
1

Hello, I’m beginner of Julia.

When I draw samples from Multivariate Normal, it uses multicores of my computers.

Especially, when N > 63, they uses 8 cores.

If this things are normal things, and If I don’t wan’t to use multicore, How can I solve this problem?

Example code is as follow.

using Distributions
using BenchmarkTools

n = 64
X=rand(n,n)
A=X’*X

@btime for i = 1 : 100
MvNormal(ones(n), A)
end

My computer is MacBook Pro 14 (M1 Pro)

2

I found a solution.

“BLAS.set_num_threads(1)”

type this sentence before sampling MvNormal.

Thank you.

1 Like
3

How do you know? I don’t observe any multi-core utilisation when running your code.

Also note that you are not drawing any samples from the distributions here, you are just instantiating the distribution objects.

d = MvNormal(ones(n), A) creates a distribution, rand(d) samples from it.

5

Oh that’s right, That code is not a sampling, but creates a distributions.

If I could upload a figure, I could explain my situation much more clearly, But I couldn’t find upload button.

But anyway, “BLAS.set_num_threads(1)” this code force my laptop only use single core.

Thank you !

6

Ah, looking at @less MvNormal(ones(n), A) I see it is doing MvNormal(ones(n), PDMat(A)) and looking at @less Distributions.PDMats.PDMat(A) I see that that is doing cholesky(A) so that’s probably where the BLAS call in the constructor comes from.

1 Like
7

Thank you!