topic PARDISO with METIS re-ordering in Intel® oneAPI Math Kernel Library

PARDISO with METIS re-ordering

Rooh — Sun, 03 Mar 2013 09:07:27 GMT

Hello:

I am using (INTEL_MKL) PARDISO solver in a visco-elasto-plastic Stokes flow code. It is a finite difference code which uses marker-in-cell technique and has a mixed formulation with linear velocity and constant pressure.

During the reordering and symbolic factorization (METIS) I keep getting SEG-faults (and occasionally a SIGBUS) at random moments, where it will probably runs through the next time. This occurs both when using METIS = 3 or 2, however it runs fine without re-ordering.

I have checked the physics, matrix conditioning etc. but the fact that solver only has a problem with longer runs, whereas it runs fine with the restart file, suggests that there is some algorithmic/compiler specific issue. I have traced the memory utilization and it is far less than the RAM.

This seems to be the problem with the INTEL_MKL version of Pardiso only.

Best regards,

Rooh

Hello Rooh.

Gennady_F_Intel — Sun, 03 Mar 2013 13:31:16 GMT

Hello Rooh.

What would happens if you will use the minimum degree algorithm?

Hi Gennady,

Rooh — Mon, 04 Mar 2013 13:57:16 GMT

Hi Gennady,

It works fine with minimum degree algorithm (iparm[1] = 0), but much slower than iparm[1] = 2 or 3

Thanks,

Rooh

Hi Rooh,

Gennady_F_Intel — Mon, 04 Mar 2013 15:07:02 GMT

Hi Rooh,

We had the problem with metis reodering in the past. what version of mkl you are using? in the case of the latest ( v.11.0 update 2), then please give us the reproducer of this problem.

--Gennady