Community
cancel
Showing results for 
Search instead for 
Did you mean: 
116 Views

intel pardiso iparm[12] or iparm(13)

Hello,

I have a question about the parameter iparm[12] (zero index, weighted matchings) of intel pardiso solver.

Actually the problem was in a solution of a non linear problem by Newton Raphson iteration. The used solver was Intel Pardiso for sparse non symmetric systems. The solution diverged, as in the topic

https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/537126

while the solver dgbsv (lapack) converged. The value of parameter was iparm[12] = 1 (as in the examples.)

I don't remember if I tried iparm[12]=0 or not (It's about a couple of months, I think I did). When I tried iparm[12] = 2 it worked and Intel Pardiso and DGBSV solvers produced the same correct result. Actually I keep using iparm[12] = 2, in difficult systems.

I cannot find the document "intel pardiso users guide" where this comment about iparm[12] = 2 was written anymore. In the site and in all MKL documents the value of iparm[12] is  0 or 1. Is it implemented the iparm[12] = 2 ? or it has the same effect as iparm[12] = 0?

In Pardiso's user guide https://pardiso-project.org/manual/manual.pdf they have a comment about iparm(13) = 2. (weighted matchings)

Thanks in advance,

 

0 Kudos
1 Reply
116 Views

Hi,

That's correct, behavior of MKL pardiso is different in case of iparm[12] is equal to 2 or 3 and sometimes results with these parameters better than 0 or 1. However the behavior currently is unstable so officially these parameters are unsupported and I recommend to not use them.

Thanks,

Alex 

Reply