topic Could you set the msglvl == 1 in Intel® oneAPI Math Kernel Library

Eigen::PardisoLLT is slower than Eigen::SimplicialLLT

Li__Jing1 — Sun, 11 Aug 2019 05:21:03 GMT

Hi,

I followed this page(https://eigen.tuxfamily.org/dox/TopicUsingIntelMKL.html) to use Eigen::PardisoLLT to solve a sparse linear system(88869x88869). I have MKL(student version) installed. I’ve successfully compiled, linked and run it. However, it’s 5 times slower than Eigen::SimplicialLLT. Am I doing anything wrong there?

Best!

Jing

What are execution times for

Gennady_F_Intel — Tue, 13 Aug 2019 03:07:58 GMT

What are execution times for Eigen::SimplicialLLT and Eigen::PardisoLLT?

For one solve, Eigen:

Li__Jing1 — Tue, 13 Aug 2019 04:20:00 GMT

For one solve, Eigen::SimplicialLLT takes 2.74s, Eigen::PardisoLLT takes 12.107s. By the way, I am solving a system Ax = B where B is a sparse matrix(88869x354)

Could you set the msglvl == 1

Gennady_F_Intel — Tue, 13 Aug 2019 05:43:56 GMT

Could you set the msglvl == 1 and share this log file?

Could you give us the reproducer?

Hi,

Li__Jing1 — Tue, 13 Aug 2019 05:59:57 GMT

Hi,

I think I've found out which part slowed the program down. I changed the right-hand side (and the variable to hold the solution) to a dense matrix. And the execution time for one solve is 1.9s. I don't know why but Eigen seemed to chop the sparse right-hand side to 4 columns and assemble the solution later, which seems to be unnecessary to me. I was expecting a larger speedup from pardiso because the CPU usage went from 20%(SimplicialLLT) to 80%(PardisoLLT with parallel MKL). I thought I can get it 4 times faster based on the CPU usage. Do you think I've got a reasonable speedup already?