Brief question on ILU and OpenMP

[Riccardo_Rossi]

Hello everyone,

i am writing here to ask about the MKL implementation of the ILU0 and ILU(k) preconditioners:

are they OpenMP parallel? if so ... only the factorization is parallel or also the "triangular" solves?

what kind of techniques is used inside? is it a multi-level ilu ...


i am asking as the parallelization of the iterative solvers is not difficult, but the parallelization of the solve phase of the ILU is highly non-trivial

thank you in advance

Riccardo

[Gennady_F_Intel]

Riccardo, both of these preconditioners are not threaded at this moment..

--Gennady

[Riccardo_Rossi]

A shame ...

is there any plan to have it parallel in future releases of MKL?

in any case thank you for the quick reply

[Gennady_F_Intel]

yes, we are working on this issue, but I cannot say in when this feature will available.

--Gennady

[red_cat]

At what stage is the development of this functions?Will the added functions for complex matrices?

[red_cat]

I am interested in - the developers read this forum?

[styc]

You can use PARDISO instead. Threaded, much stronger than ILU, has built-in CGS.

[Riccardo_Rossi]

Hello,
this is just to say that pardiso works very nicely and is very fast ... but it is a DIRECT SOLVER!!

ILU is for solving really large systems, which can not be solved directly.
I do understand that it is extremely difficult to write an efficient ILU preconditioner, that's exactly the reason for which i orginally started this thread

Riccardo

[Gennady_F_Intel]

Hello Riccardo,

1) regarding the original question -we do nothave plans at this time to add this feature to the next release of Intel MKL.

2) qoute: "ILU is for solving really large systems, which can not be solved directly". Riccardo, how about out-of-core version of PARDISO? usage of this mode allow to solve the real big task.

--Gennady

[red_cat]

I want to use iterative methods (Bi-CG Stab, TFQMR and other) for solving dense linear system. I'm doing at the beginning prefiltration dense matrix to obtain a sparse preconditioner, and then use ILU0 or ILUT algorithm for decomposition.

[Schaefer__Florian]

Since the last post on this topic has been a couple of years ago, I'd like to ask, what is the current state of parallelism for ilu0 preconditioners? Is such a feature on the horizon (or implemented already?).