when using pardiso on an upper triangular symmetric matrix of dimension 130,00 x 130,000 with 1,000,000 non-zeros it needs about 15 sec for a factorization, which is very slow compared to cholmod(suitesparse) 0.3 sec.
Further on large matrices, say >1,000,000 x >1,000,000 pardiso is so unbearable slow that it becomes unusable, whereas cholmod(suitesparse) can easily cope with the problem.
Is there any way to speed up pardiso? I figured that it gets stuck at the analysis stage.
I am using oneapi version 2021.2.0 on linux, kernel version 5.16.
The only way to speed up the reordering stage is by trying the OpenMP version of Metis (iparm=3).
wrt to factorization performance against cholmod -- here make sense to share the case to check the behavior on our end.
You could upload the case anywhere ( google drive as an example) and give us the link by the private channel into this forum or directly via email ( Gennady.Fedorov@intel.com). And please add the driver which will write this matrix and run the solver with your parameters. It will essentially save the reproducing time.
I have emailed you a link for downloading an archive with matrix and driver.
The example also reproduces the observation from this thread https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/omp-warning-when-using-pardiso-and-metis-threaded-version/td-p/1363562
Thanks for looking into this.
I responded by email to you on this matter.
The reordering performance problem has been reproduced on our end and escalated against the Sparse Solver Team.
We will keep this thread updated in case of any updates.
I had contacted the pardiso team about this problem as well, and according to their test and a test library they provided to me the problem does not affect the most recent pardiso version.
Does Intel have any plans for updating pardiso?
The issue has already been escalated to the mkl team.
Due to many changes, we might not have a solution at this time.
I will consult with the mkl team and will get back to with an estimated timelione.