topic Thank you Irina. in Intel® oneAPI Math Kernel Library

Using Pardiso to solve linear system with the same pattern

Hwang__Carpenter — Wed, 13 Dec 2017 20:34:12 GMT

Suppose that I need to solve a linear system in this form, Ax=b, in many iterations. In each iteration, I get a new matrix A. However, the pattern of the matrix is the same. That is, the indices of the nonzeros in A are the same throughout the iterations, but the values of the nonzeros change throughout the iterations.

I don't want to go from phase 1 to phase 3 every time. Since LU factorization doesn't depend entirely on the values of the nonzeros, can I just run phases 1 and 2 just once and in the iterations run phase 3 for different IA, JA, and A vectors, so that I don't have to factorize every time?

Hello,

Irina_S_Intel — Thu, 14 Dec 2017 21:37:00 GMT

Hello,

I think what you need is a low rank update - it allows to use factorization for matrix A1 when factorizing A2 if both matrices have the same pattern of nonzeroes.

To use the low rank update feature, set iparm(39) = 1 and iparm(24) = 10. And additionally supply an array that lists the values in A2 that are different from A1 using the perm parameter.

More information here.

Best regards,

Irina

Thank you Irina.

Hwang__Carpenter — Sun, 17 Dec 2017 03:40:02 GMT

Thank you Irina.

Should I release memory each time (i.e. using Phase = -1)? As you know, I'm solving a linear system in many iterations. If I don't release memory, will my code accumulate memory usage unnecessarily?