Since julia-1.10 / OpenBLAS32-0.3.22 the number of threads must be explicitly specified

[fp4code]

Since julia-1.10 a simple program like this one is very slow because only one CPU is used

# On a 32-core computer, uncomment next line to have 3200% CPU workload if julia version is > 1.9
# ENV["OPENBLAS_NUM_THREADS"]="32"

using Random, MUMPS, MPI, SparseArrays, LinearAlgebra
N = 10000
Nc = 1
Random.seed!(3)
A = sprand(N, N, 0.1) + I
rhs = rand(N, Nc)
x = Matrix{Float64}(undef, N, Nc)
MPI.Init()
mumps = Mumps{Float64}(mumps_unsymmetric, default_icntl, default_cntl64)
associate_matrix!(mumps, A)
t = @elapsed factorize!(mumps)
associate_rhs!(mumps, rhs)
solve!(mumps)
MUMPS.get_sol!(x, mumps)
finalize(mumps)
MPI.Finalize()
println("maximum error = ", maximum(abs.(A*x - rhs)), ", factorise time (s) = ", t)

No such problem with pinning OpenBLAS32_jll to 0.3.21: Pkg.add(name="OpenBLAS32_jll", version="0.3.21")

One solution is to set the environment variable for the desired number of threads: export OPENBLAS_NUM_THREADS=32

This can be done inside Julia: ENV["OPENBLAS_NUM_THREADS"]="32"

[amontoison]

@ViralBShah Are you aware of this issue with the last release of OpenBLAS?

[fp4code]

@ViralBShah @amontoison There was a typo in the title (now corrected), the problematic package is OpenBLAS32_jll, not OpenBLAS_jll

Same problem with the last release of OpenBLAS32_jll (0.3.25):

Slow with default configuration:

Pkg.add(url="https://github.com/JuliaBinaryWrappers/OpenBLAS32_jll.jl", rev="main")
Pkg.add(name="OpenBLAS32_jll", version="0.3.24")
Pkg.add(name="OpenBLAS32_jll", version="0.3.23")
Pkg.add(name="OpenBLAS32_jll", version="0.3.22")

Fast with default configuration:

Pkg.add(name="OpenBLAS32_jll", version="0.3.21")

[guiburon]

FYI, in my case I have to set the env variable OMP_NUM_THREADS to the desired number of threads per proc.

OPENBLAS_NUM_THREADS set unlock multithreading per proc but its value seems irrelevant: all the CPUs of my system are used. This can lead to slow computation with too many threads.

[ViralBShah]

I suspect this is because with OpenBLAS, we set the number of threads in our __init__ when LinearAlgebra loads (IIRC), but we don't do such a thing with OpenBLAS32_jll which is a BB generated package.

I had created OpenBLAS32.jl, which may be a place where we can specify number of threads on loading and such.

[amontoison]

@ViralBShah
A little bit off-topic but you should check in OpenBLAS32.jl that an LP64 BLAS / LAPACK is not already loaded.
I do that for all packages compiled LBT, like Ipopt.jl or HSL.jl.
A user could use MKL or AppleAccelerate already and you don't to want to modify the setup.
https://github.com/JuliaSmoothOptimizers/HSL.jl/blob/main/src/HSL.jl#L16-L23

[amontoison]

@fp4code
I think that I found what we need:
https://github.com/JuliaLang/julia/blob/master/stdlib/LinearAlgebra/src/LinearAlgebra.jl#L766

Can you test in MUMPS.jl with an __init__ function?