topic Hi, in Intel® oneAPI Math Kernel Library

Pardiso slow performance with multiple RHS

vahid_s_ — Fri, 06 Dec 2013 02:50:47 GMT

I am solving K*D=F equation with pardiso sparse solver. K is a 11000*11000 sparse symmetric matrix and F is also sparse matrix with 11000*363 dimensions. On a dell computer with Intel Core 2 Duo CPU and MKL version 10.0.1 it takes 2.2 seconds to solve the equations. I am using default parameters for the solver and I tried both sequentional and multi-threading options but I can not improve the performance time. I found that the performance is very sensitive to the number of RHS:

Number of RHS=363, Time=2.21 seconds

Number of RHS=120, Time=0.53 seconds

Number of RHS=30, Time=0.19 seconds

Number of RHS=9, Time=0.06 seconds

Why there is much difference in performance?! In my program number of RHS can be between 1-1000 for different situations. Is there any way to speed up the process for high number of RHS?

If I switch to the new version of MKL can I get significant improvement?

Thank you in advance for your help.

yes, please try the latest 11

Gennady_F_Intel — Fri, 06 Dec 2013 03:04:19 GMT

yes, please try the latest 11.1. you can get eval version for 30 days.

Hi,

vahid_s_ — Fri, 13 Dec 2013 21:33:49 GMT

Hi,

I tried the latest version. It gives me better performance but the improvement is not that much! for example for Number of RHS=363:

Version 10.1: Time=2.21 seconds

Version 11.1: Time=1.70 seconds

I made a more detailed topic about my question (MKL: Partial factorization and Schur Complement?) hope the answers from that help me to improve the performance.

Thank you,