I am comparing pardiso performance on matrix (size 2500 rows with about 8000 non-zeros). The matrix is concentrated with entries clustered around the diagonal with some values sprinkled sparsely.
I am comparing performance with Kundert's sparse matrix (used for circuit simulation software SPICE http://sparse.sourceforge.net/). The performance of PARDISO is more than a order of magnitude WORSE than Kunderts.
I wanted to know if this is expected or if there are tuning that I can do to PARDISO input to take advantage of the matrix structure
I can post the data files and code if needed. Thanks
No sure if it is still interested. Just for your information, MKL have new version released successively. there are some improvement about PARDISO, Please see
What's New in Intel® MKL 11.0 update 2
- Introduced Intel MKL Extended Eigensolver:
Intel MKL Extended Eigensolver is a high performance package for solving symmetric standard or generalized symmetric-definite eigenvalue problems on matrices in dense, LAPACK banded, and sparse (CSR) formats. It is based on an innovative fast and stable numerical algorithm named Feast (See Attributions section below)
What's New in Intel® MKL 11.0 update 1
- PARDISO: Imaginary part of the diagonal values for Hermitian matrices are ignored