[Trilinos-Users] [EXTERNAL] Re: AMG Preconditioner Neumann Laplacian 32 vs 64 bit issue

Fri Dec 2 00:55:47 MST 2011

 Hi Martin,

 I guess that this funky thing Ray is referring
 to is highly sensitive to rounding errors, which
 in turn depend on the compiler/architecture. This may
 explain why you are observing a different behaviour
 among both machines.

 Regards,
  Alberto.

 On Thu, 1 Dec 2011 21:42:18 +0100, Martin Stoll wrote:
> Hi Ray,
>
> thanks for this. It is indeed very useful to know, I guess I should
> have known already.
>
> Do you see any reason why my 32bit machine gets away with the same
> setup?
>
> Martin
>
>  On Thu, Dec 1, 2011 at 9:28 PM, Tuminaro, Raymond S  wrote:
>
>> My guess is that you have understood the issue ... but just in case
>> it was not totally clear. ML's interpolation
>> is exact for constants. This means that the coarsest level matrix
>> is still singular. We then use a direct solver
>> for this system (by default). The direct solver might do something
>> funky with the singular system, and
>> then ML will happily interpolate this funky thing up the
>> hierarchy. ML often works with singular systems, but
>> it is always a bit scary and it is best to just do an iterative
>> solve for the coarsest level system.
>> -Ray
>>
>> -------------------------
>>
>> FROM: Martin Stoll [martin.stoll80 at gmail.com [18]]
>> SENT: Thursday, December 01, 2011 12:16 PM
>> TO: Tuminaro, Raymond S
>> CC: Alberto F. Martín-Huertas; trilinos-users at software.sandia.gov
>> [19]
>> SUBJECT: Re: [EXTERNAL] Re: [Trilinos-Users] AMG Preconditioner
>> Neumann Laplacian 32 vs 64 bit issue
>>
>> Dear Ray,
>>
>> thanks for the comments.
>>
>> Yes, the Neumann Laplacian is indeed singular. In fact, from
>> within dealII I simply use the ML AMG preconditioner as a black box
>> as there are only a few parameters I can play with. In my case
>> usually, type of smoother, smoother steps and number of cycles. So
> I
>> am not really aware what is happening within ML and I go with all
>> the default settings. I am sorry for being so ignorant but I never
>> had any problems. Only when I moved my code to a different machine
>> running 64 bits I was having problems. Again, same version of
> dealII
>> and trilinos on my mac runs smoothly.
>>
>> I will look into changing some options from within deal.II to try
>> a different solver for the coarse grid and hope for the best.
>>
>> Thanks again for the comments.
>>
>> Best,
>> Martin
>>
>> 2011/12/1 Tuminaro, Raymond S
>>
>>> So, I'm guessing that the Neumann Laplacian is singular.  Are
>>> you using a direct solve on the coarsest grid? Depending on how
>>> things are handled here ... you might get different results. You
>>> might try switching to an
>>> iterative method on the coarsest level and see what happens.
>>>
>>> -Ray
>>>
>>> -------------------------
>>>
>>> FROM: trilinos-users-bounces at software.sandia.gov [10]
>>> [trilinos-users-bounces at software.sandia.gov [11]] on behalf of
>>> Martin Stoll [martin.stoll80 at gmail.com [12]]
>>> SENT: Thursday, December 01, 2011 4:29 AM
>>> TO: Alberto F. Martín-Huertas
>>> CC: trilinos-users at software.sandia.gov [13]
>>> SUBJECT: [EXTERNAL] Re: [Trilinos-Users] AMG Preconditioner
>>> Neumann Laplacian 32 vs 64 bit issue
>>>
>>> Hi Alberto,
>>> thanks for your comment.
>>>
>>> I just tried the Dirichlet Laplacian on  the velocity space,
>>> the Neumann Laplacian was on the pressure space, and that worked
>>> fine.
>>>
>>> As I said on my laptop and the 32 bit desktop everything work
>>> but when I switch to the 64 bit machine the outcome is quite
>>> different. Our software engineer also tried different versions of
>>> trilinos (currently 10.4.2, which is recommended by the deal.II
>>> developers). My Desktop uses trilinos 9.03 and my laptop 10.4.2
>>> and they both work fine.
>>>
>>> Any help is greatly appreciated.
>>>
>>> Best,
>>> Martin
>>>
>>> 2011/12/1 Alberto F. Martín-Huertas
>>>
>>>>  Hi Martin,
>>>>
>>>>  did you observe the same behaviour with a
>>>>  Dirichlet Laplacian ?
>>>>
>>>>  Best regards,
>>>>  Alberto.
>>>>
>>>>  On Thu, 1 Dec 2011 12:52:08 +0100, Martin Stoll wrote:
>>>> > Hi all,
>>>> >
>>>> > I am using deal.II and from there call the ML AMG
>>>> preconditioner. I
>>>> > am using it to approximate a Neumann Laplacian, which works
>>>> like a
>>>> > charm on
>>>> > my 32 bit desktop Ubuntu machine and also on my Mac Book Pro
>>>> running
>>>> > Snow Leopard. Now copying this to a linux 64 bit machine I
>>>> get weird
>>>> > results meaning
>>>> >  that applying the preconditioner results
>>>> > 32 bit
>>>> > norm(vin)=0.01
>>>> > APPLY AMG
>>>> > norm(vout)=0.15
>>>> >
>>>> > 64 bit
>>>> > norm(vin)=0.01
>>>> >  APPLY AMG
>>>> >  norm(vout)=1e+11
>>>> >
>>>> > When instead of the Neumann Laplacian I put a mass matrix
>>>> into the
>>>> > AMG routine the blow does not happen.
>>>> >
>>>> > I am not sure where to go from here so I hope I could get
>>>> some help
>>>> > here.
>>>> >
>>>> > Best,
>>>> >  Martin
>>>> >
>>>> > --
>>>> > Martin Stoll
>>>> > Postdoctoral Research Fellow
>>>> >
>>>> > Computational Methods in Systems and Control Theory
>>>> >
>>>> > Max Planck Institute for Dynamics of Complex Technical
>>>> Systems
>>>> > Sandtorstr. 1, D-39106 Magdeburg,Germany
>>>> > Email: martin.stoll80 at gmail.com [1] [1]
>>>> > URL: http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/ [2] [2]
>>>>
>>>>>
>>>> > Tel :+49 391 6110 384 [3]
>>>> >
>>>> >
>>>> >
>>>> > Links:
>>>> > ------
>>>> > [1] mailto:stollm at mpi-magdeburg.mpg.de [4]
>>>> > [2] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/ [5]
>>>>
>>>> --
>>>>  Alberto F. Martín-Huertas
>>>>  Centre Internacional de Mètodes Numèrics a l'Enginyeria
>>>> (CIMNE)
>>>>  Parc Mediterrani de la Tecnologia, UPC
>>>>  Esteve Terrades 5, Building C3, Office 210,
>>>>  08860 Castelldefels (Barcelona, Spain)
>>>>  Tel.: (+34) 9341 34223 [6]
>>>>  e-mail: amartin at cimne.upc.edu [7]
>>>>
>>>>  ________________
>>>>  IMPORTANT NOTICE
>>>>  All personal data contained on this mail will be processed
>>>>  confidentially and registered in a file property of CIMNE in
>>>>  order to manage corporate communications. You may exercise
>>>> the rights
>>>>  of access, rectification, erasure and object by
>>>>  letter sent to Ed. C1 Campus Norte UPC. Gran Capitán s/n
>>>> Barcelona.
>>>>
>>>> _______________________________________________
>>>> Trilinos-Users mailing list
>>>> Trilinos-Users at software.sandia.gov [8]
>>>> http://software.sandia.gov/mailman/listinfo/trilinos-users [9]
>>>
>>> --
>>> Martin Stoll
>>> Postdoctoral Research Fellow
>>>
>>> Computational Methods in Systems and Control Theory
>>> Max Planck Institute for Dynamics of Complex Technical Systems
>>> Sandtorstr. 1, D-39106 Magdeburg,Germany
>>>
>>> Email: martin.stoll80 at gmail.com [15]
>>>
>>> URL: http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/ [16]
>>>
>>> Tel :+49 391 6110 384 [17]
>>
>> --
>> Martin Stoll
>> Postdoctoral Research Fellow
>>
>> Computational Methods in Systems and Control Theory
>> Max Planck Institute for Dynamics of Complex Technical Systems
>> Sandtorstr. 1, D-39106 Magdeburg,Germany
>>
>> Email: martin.stoll80 at gmail.com [21]
>>
>> URL: http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/ [22]
>>
>> Tel :+49 391 6110 384 [23]
>
> --
> Martin Stoll
> Postdoctoral Research Fellow
>
> Computational Methods in Systems and Control Theory
> Max Planck Institute for Dynamics of Complex Technical Systems
> Sandtorstr. 1, D-39106 Magdeburg,Germany
> Email: martin.stoll80 at gmail.com [25]
>
> URL: http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/ [26]
>
> Tel :+49 391 6110 384
>
>
>
> Links:
> ------
> [1] mailto:martin.stoll80 at gmail.com
> [2] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/
> [3] http://webmail.cimne.upc.edu/tel:%2B49%20391%206110%20384
> [4] mailto:stollm at mpi-magdeburg.mpg.de
> [5] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/
> [6] http://webmail.cimne.upc.edu/tel:%28%2B34%29%209341%2034223
> [7] mailto:amartin at cimne.upc.edu
> [8] mailto:Trilinos-Users at software.sandia.gov
> [9] http://software.sandia.gov/mailman/listinfo/trilinos-users
> [10] mailto:trilinos-users-bounces at software.sandia.gov
> [11] mailto:trilinos-users-bounces at software.sandia.gov
> [12] mailto:martin.stoll80 at gmail.com
> [13] mailto:trilinos-users at software.sandia.gov
> [14] mailto:amartin at cimne.upc.edu
> [15] mailto:stollm at mpi-magdeburg.mpg.de
> [16] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/
> [17] http://webmail.cimne.upc.edu/tel:%2B49%20391%206110%20384
> [18] mailto:martin.stoll80 at gmail.com
> [19] mailto:trilinos-users at software.sandia.gov
> [20] mailto:rstumin at sandia.gov
> [21] mailto:stollm at mpi-magdeburg.mpg.de
> [22] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/
> [23] http://webmail.cimne.upc.edu/tel:%2B49%20391%206110%20384
> [24] mailto:rstumin at sandia.gov
> [25] mailto:stollm at mpi-magdeburg.mpg.de
> [26] http://www.mpi-magdeburg.mpg.de/mpcsc/stollm/

-- 
 Alberto F. Martín-Huertas
 Centre Internacional de Mètodes Numèrics a l'Enginyeria (CIMNE)
 Parc Mediterrani de la Tecnologia, UPC
 Esteve Terrades 5, Building C3, Office 210,
 08860 Castelldefels (Barcelona, Spain)
 Tel.: (+34) 9341 34223
 e-mail: amartin at cimne.upc.edu

 ________________
 IMPORTANT NOTICE
 All personal data contained on this mail will be processed 
 confidentially and registered in a file property of CIMNE in
 order to manage corporate communications. You may exercise the rights 
 of access, rectification, erasure and object by
 letter sent to Ed. C1 Campus Norte UPC. Gran Capitán s/n Barcelona.