topic Re: Using Colamd with Pardiso in Intel® oneAPI Math Kernel Library

Using Colamd with Pardiso

damienhocking — Wed, 16 Sep 2009 01:57:04 GMT

Hi all,

I was wondering if anyone has used Colamd with Pardiso. Given that colamd is a column reordering and Pardiso is a row-ordered solver this might be slow because the matrix would have to be reordered into column form. Anyone tried this? I'm trying Scotch out right now as well.

Damien

Re: Using Colamd with Pardiso

ArturGuzik — Wed, 16 Sep 2009 03:52:46 GMT

Quoting - damienhocking

Pardiso is a row-ordered solver this might be slow because the matrix would have to be reordered into column form.

Damien,

what do you mean by that?

A.

Re: Using Colamd with Pardiso

damienhocking — Thu, 17 Sep 2009 00:42:51 GMT

Quoting - ArturGuzik

Damien,

what do you mean by that?

A.

Pardiso requires the matrix in row-major format. Colamd requires the matrix in column-major format. There would need to be a reordering pass and a separate structure stored to map the reordering back to row-major.

This is starting to look like more CPU work than it could be worth.

Re: Using Colamd with Pardiso

Gennady_F_Intel — Mon, 21 Sep 2009 05:05:32 GMT

Quoting - damienhocking

Damien,

Yes, it will require more CPU work, however we don't expect the visible slowdown of PARDISO.

The slowdown will be negligible especially for the big problem sizes.

--Gennady

Re: Using Colamd with Pardiso

damienhocking — Wed, 14 Oct 2009 17:33:58 GMT

Quoting - Gennady Fedorov (Intel)

Damien,

Yes, it will require more CPU work, however we don't expect the visible slowdown of PARDISO.

The slowdown will be negligible especially for the big problem sizes.
--Gennady

I have a client with some difficult and badly-scaled matrices from their simulation environment. With some careful scaling in my nonlinear solver, the parallel METIS ordering is doing a good enough job and Pardiso solves everything now with good speedups on multicores.

Damien