[Trilinos-Users] [EXTERNAL] Re: Multigrid preconditioning with 64 bit support

[Chris Jackson]

Hi Ray,

Thanks for this information.  

1) Yes, it is most often due to mesh stretching.
2) It is finite volume.

Regards

Chris

-----Original Message-----
From: Raymond Tuminaro [mailto:rstumin at sandia.gov] 
Sent: Thursday, March 07, 2013 5:05 PM
To: Chris Jackson
Cc: trilinos-users at software.sandia.gov
Subject: Re: [Trilinos-Users] [EXTERNAL] Re: Multigrid preconditioning with 64 bit support

Chris,
    This answer to this question is a bit tricky. The basic tool that ML has to detail with anisotropic problems is that of dropping weak connections. This can be controlled by
           MLList.set("aggregation: threshold", t);

where t is some double precision number usually between 0 and .2 I don't have great advice for picking this value ... but roughly it should be chosen in the following way. Consider first symmetrically scaling the matrix so that it has ones on the diagonal. That is
             Ahat = D A D
where D is a diagonal matrix and D_ii = 1/sqrt(A_ii). Then we would want to pick t so that it is smaller than any abs(Ahat_ij)  where this nonzero corresponds to a strong direction, but larger than any abs(Ahat_ij) which corresponds to a weak direction. Does this sort of make sense?

I should ask two other questions:
     1) is the anisotropic behavior due to mesh stretching (as opposed to within the PDE operator)?
     2) how are you discretizing your system, finite elements, finite differences, finite volumes?

The main reason I ask this is that finite elements tend to smear coefficients and so picking an appropriate t is quite problematic. If you are using finite elements and your anisotropic behavior is due to the mesh, then there are a couple of other ML tricks that we can discuss to deal with this.

-Ray


On 03/07/13 10:33, Chris Jackson wrote:
> As an aside, we use ML as a preconditioner for BiCGStab. With that 
> method, we have very slow convergence rates on strongly anisotropic 
> Poisson problems, when compared to more isotropic ones. If anyone knows of any special settings for ML to address that, I'd be keen to know.
>
> Regards,
>
> Chris Jackson
>
> *From:*trilinos-users-bounces at software.sandia.gov 
> [mailto:trilinos-users-bounces at software.sandia.gov] *On Behalf Of 
> *C.S. Natarajan
> *Sent:* Thursday, March 07, 2013 9:44 AM
> *To:* trilinos-users at software.sandia.gov
> *Subject:* Re: [Trilinos-Users] [EXTERNAL] Re: Multigrid 
> preconditioning with 64 bit support
>
> Thanks Mike.
>
> On Thu, Mar 7, 2013 at 7:43 AM, Heroux, Michael A <maherou at sandia.gov <mailto:maherou at sandia.gov>> wrote:
>
> ML does not support the "long long" map features in Epetra. We do not 
> plan to extent ML to support this feature, since Muelu is coming along and natively supports 64-bit types.
>
> Mike
>
> *From: *"C.S. Natarajan" <csnataraj at gmail.com 
> <mailto:csnataraj at gmail.com>>
> *Date: *Wednesday, March 6, 2013 11:31 PM
> *To: *Eric Marttila <eric.marttila at thermoanalytics.com 
> <mailto:eric.marttila at thermoanalytics.com>>
> *Cc: *"trilinos-users at software.sandia.gov <mailto:trilinos-users at software.sandia.gov>"
> <trilinos-users at software.sandia.gov 
> <mailto:trilinos-users at software.sandia.gov>>
> *Subject: *[EXTERNAL] Re: [Trilinos-Users] Multigrid preconditioning 
> with 64 bit support
>
> Hi Eric,
> I don't have a problem building ML on a 64 bit platform. I am trying 
> to solve a large problem (strongly anisotropic
> poisson) and was wondering if ML has 64-bit support. Just to clarify, 
> have you used ML in conjunction with long long Epetra_Map's?
> I have no problem with ML up to Trilinos 10.10 but I just upgraded to 
> 11.03 to solve bigger problems and have started getting errors when building the preconditioner.
>
> Thanks for the help.
>
> C.S.N
>
> On Wed, Mar 6, 2013 at 10:48 PM, Eric Marttila 
> <eric.marttila at thermoanalytics.com
> <mailto:eric.marttila at thermoanalytics.com>> wrote:
>
> Hi,
> I'm not sure exactly what kind of 64bit support you're looking for, 
> but I've built and used the ML package on 64-bit linux without any 
> issues.
> --Eric
>
>
> On Wednesday, March 06, 2013 05:36:58 pm C.S. Natarajan wrote:
>>  Hello,
>>  Is there a multigrid preconditioning package with 64  bit support? 
>> It looks like ML is 32-bit. I am using  trilinos11.0.3.
>>
>>  Thanks,
>>  C.S.N
>
>
>
> This body part will be downloaded on demand.

-- 
Ray Tuminaro                      phone: (925) 294-2564
MS 9159                           fax:   (925) 294-2234
Sandia National Laboratories      email: rstumin at sandia.gov
PO Box 969                        http://www.cs.sandia.gov/~rstumin
Livermore, CA 94551