Cluster sparse solver crash after switching to oneMKL 2025.0 (windows, 64-bit indices)

[DmitrySmi]

Hello Intel Developers,

We just switched to oneMKL 2025.0 and found some of our numerical simulation code crashing. Enclosed to this post is a simple reproducer, which worked fine in 2021.3, but crashes in 2025.0. The reproducer factorizes complex symmetric indefinite matrix. The symbolic analysis seems to work fine, but the subsequent cluster_sparse_solver call with phase = 22 (for numerical factorization) crashes. The platform is windows, and we build it with 64-bit indexing (ilp64)

Thanks in advance for looking!

Dmitry

[Fengrui]

Hi Dmitry,

Thank you for reaching out to us!

I can reproduce the crash at numerical factorization when using 1 MPI process. In this case, turning on msglvl, it shows that actually oneMKL SMP Pardiso solver is called.

However, if I use more than 1 MPI processes, I got through all the three stages, and obtained the solution,

The solution of the system is:
 x [0] = 0.095477 + j*0.000000
 x [1] = 0.045226 + j*0.000000
Relative residual = 7.850462e-17

In this case (np>1), I got errors when trying to get the size of L factor,

Getting size and numbers of nonzeros for the factors L and U... Export functionality error: CODE PATH = 1 not supported.

ERROR during getting sizes for L data: -9

It's likely something wrong in the solver. I will get back to you later.

Thanks,

Fengrui

[DmitrySmi]

Fengrui, thanks for getting back to me.

Here is one more detail to add -- we found that if we specify the solver parameters explicitly ( iparm[0] = 1 and others ) instead of letting solver use the defaults ( iparm[0] = 0 ) the crash goes away.

So we have a good working workaround for this problem now.

[Fengrui]

Hi Dmitry, the team is currently working on this issue. 

As for the workaround to set parameters explicitly, I observed an error at symbolic factorization when trying to use 2 MPI processes to run this simple test (passed if only using 1). Do you also see this behavior for the reproducer/real cases?

Thanks,

Fengrui 

[DmitrySmi]

Hi Fengrui,

We only use it with one process, so I can not comment on multi-process behavior.

Our only reason for using the cluster sparse solver (instead of using PARDISO directly) is that we need access to exported L and U factors (reasons for that outlined in another thread here MKL PARDISO, get L U factors - Intel Community)

Thanks

Dmitry

[Fengrui · Accepted solution]

Hi Dmitry,

The issue has been fixed. The fix will be included in the 2025.2 release.

Thanks,

Fengrui

[DmitrySmi]

Thanks much!

[Fengrui]

Hi Dmitry,

Did you get a chance to try the latest 2025.2 release?

Best,

Fengrui

[DmitrySmi]

Hi Fengrui -- not yet unfortunately, but I hope to do it soon. Thanks for the fix!