[Trilinos-Users] [EXTERNAL] ilut

Bradley, Andrew Michael ambradl at sandia.gov
Fri Jan 22 13:59:43 EST 2016


Hi Danielle,


OK, these are good checks.


I think if I were debugging this, I would do two high-level things next.


First, make a small, well-conditioned problem. Then run it with 1 to N MPI ranks and observe the results. Since the matrix is small by construction, one can also print the entries from each processor. (I'm a bit busy, so I can't help with the details of printing entries; sorry.)


Second, increase the quality of the preconditioner. Do this in two ways: increase the subdomain overlap (AZ_overlap) to 2 or 3, and possibly also make the drop tolerance within each domain smaller. See if this resolves the 3-processor case. I'm not suggesting we want to have to do this in practice at large scale, but in this debug stage, these are useful knobs to help with debugging.


Andrew


________________________________
From: Danielle Catherine Maddix <dcmaddix at stanford.edu>
Sent: Friday, January 22, 2016 11:48 AM
To: Bradley, Andrew Michael; trilinos-users at trilinos.org
Subject: Re: [EXTERNAL] [Trilinos-Users] ilut


Hi Andrew,

Thanks for the help!


1. With no preconditioner, 1,2,3 and 4 processors and no preconditioner, I get the same answer.  It does not converge to the correct solution due to ill-conditioning. but is consistent amongst the processors.

2. Yes, the 12,4-proc cases all give the same correct solution with the ilut preconditioner.

Let me know what you think.


Thanks,
Danielle


________________________________
From: Bradley, Andrew Michael <ambradl at sandia.gov>
Sent: Friday, January 22, 2016 10:24 AM
To: Danielle Catherine Maddix; trilinos-users at trilinos.org
Subject: Re: [EXTERNAL] [Trilinos-Users] ilut


Hi Danielle,


1. What happens if you do not use a preconditioner at all?


2. Have you verified in the 1,2,4-proc cases that the output solutions all agree?


Cheers,

Andrew


________________________________
From: Trilinos-Users <trilinos-users-bounces at trilinos.org> on behalf of Danielle Catherine Maddix <dcmaddix at stanford.edu>
Sent: Friday, January 22, 2016 11:20 AM
To: trilinos-users at trilinos.org
Subject: [EXTERNAL] [Trilinos-Users] ilut


Hello,


I am solving a FSI finite element matrix vector system.  I am using GMRES with glut in AztecOO.  It appropriately solves it for 1,2 and 4 processors on my local machine, which with 3 processors it cannot bring the residual down below 1e-1.  I was wondering if this could be an issue with the ilut preconditioner, since it says convergence depends on the domain decomposition, or if it is possible there is a bug in my code?



Thanks,

Danielle
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://trilinos.org/pipermail/trilinos-users/attachments/20160122/26620c40/attachment.html>


More information about the Trilinos-Users mailing list