Incomplete Cholesky decomposition

Robin_T_ — Fri, 29 Nov 2019 14:33:43 GMT

Hi,

my question relates to this previous thread: https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/550882

I am trying to solve a large (N > 50e6) sparse (~1e-6) symmetric linear system of equations.

With PARDISO I am running into memory problems with a 576 GB RAM machine, even when utilizing the OOC capability.

So, I am trying to implement the RCI CG method. I need a suitable preconditioner because I am very quickly reaching large numbers of iterations with the CG method. How can I implement an incomplete Cholesky factorization? MKL only provides LU factorization apparently, which could be used in conjunction with GMRES. However, then I would have to provide the whole matrix (not just one triangle) in CSR format. It might be easier to apply incomplete Cholesky factorization. Is there any way to do so?

Thank you for any advice.

Re: Incomplete Cholesky decomposition

Royi — Fri, 10 Jul 2020 07:34:59 GMT

I also wait for Incomplete Cholesky Decomposition support in Intel MK.

I hope they will support both Thresholding and Zero Filling with Modified Incomplete Cholesky Decomposition.

I also expects it to use modern techniques so it will be parallel and the fastest algorithms out there.

topic Re: Incomplete Cholesky decomposition in Intel® oneAPI Math Kernel Library

Incomplete Cholesky decomposition

Re: Incomplete Cholesky decomposition