Pardiso.jl
Pardiso.jl copied to clipboard
JuliaSparse
Failure on Intel Mac
I've encountered this error while running Pardiso on an Intel Mac. (I observed this largely on Julia v1.7 but I could not get a chance to double check whether this happens on 1.6). It causes the test failure here
The error is like:
Raster Pairwise: Error During Test at /Users/runner/work/Circuitscape.jl/Circuitscape.jl/src/utils.jl:308
Got exception outside of a @test
Not enough memory.
Stacktrace:
[1] check_error(ps::Pardiso.MKLPardisoSolver, err::Int32)
@ Pardiso ~/.julia/packages/Pardiso/3uj3F/src/mkl_pardiso.jl:79
[2] ccall_pardiso(ps::Pardiso.MKLPardisoSolver, N::Int64, nzval::Vector{Float64}, colptr::Vector{Int64}, rowval::Vector{Int64}, NRHS::Int64, B::Vector{Float64}, X::Vector{Float64})
@ Pardiso ~/.julia/packages/Pardiso/3uj3F/src/mkl_pardiso.jl:73
[3] pardiso(ps::Pardiso.MKLPardisoSolver, X::Vector{Float64}, A::SparseArrays.SparseMatrixCSC{Float64, Int64}, B::Vector{Float64})
@ Pardiso ~/.julia/packages/Pardiso/3uj3F/src/Pardiso.jl:346
[4] #_#36
@ ~/work/Circuitscape.jl/Circuitscape.jl/src/utils.jl:459 [inlined]
(Full error)[https://github.com/Circuitscape/Circuitscape.jl/runs/4644991065?check_suite_focus=true#step:6:180]
To reproduce:
using JLD2, FileIO
ps = MKLPardisoSolver()
A = load("mat2.jld2")["G"]
B = rand(size(A,1))
X = zeros(size(A,1))
solve!(ps, X, A, B)
I've attached the matrix JLD2 here.
This example works on Julia v1.6.5.
That's surprising to me. I don't see how a different Julia version would influence this.
Not failing on my intel mac with 1.7.3 - but I am getting incorrect answers. But circuitscape tests on trivial sized matrices seem to be running out of memory and complaining about ordering (not the case in 1.6).
https://github.com/Circuitscape/Circuitscape.jl/runs/5368204337?check_suite_focus=true
I am hoping that MKL 2022 may be somehow better. Opened a PR for that.
Out of memory errors and complaint about matrix ordering are also symptomatic of the problems in #73 and #88. It could be that the error is caused by a 32bit vs 64bit error when calling MKL. The issue appear sporadically since it depends on the load order of other packages, which might explain why it is not easily reproducible.
