[Trilinos-Users] AztecOO conjugate gradient

C.S. Natarajan csnataraj at gmail.com
Thu May 9 16:04:31 MDT 2013


Hi Bart,
           At the face of it, 15s doesn't sound bad. I would think that
Gigabit ethernet probably doesn't matter for now as the operations are
memory bandwidth bound. Sorry, I am not sure I uderstand what 5594877 means

If this problem is a purely isotropic poisson problem I am a little
surprised that SGS is blowing up. SGS algorithm is essential local
processor based, so I would think that it makes sense that the solun
/convergence rate would depend on nproc but I dont know why it would blow
up. Maybe one of the Trilinos developers can clarify on this issue?

Chebyshev/Polynomial smoothers are definitely a solid option depending on
the problem. In case you dont have this already, here is a paper by some of
the main developers of ML. http://www.columbia.edu/~ma2325/adams_poly.pdf.

Cheers,
C.S.N
On Thu, May 9, 2013 at 2:09 PM, Bart Janssens <bart at bartjanssens.org> wrote:

>  On Thu, May 9, 2013 at 6:33 PM, C.S. Natarajan <csnataraj at gmail.com>wrote:
>
>>                Bart I have run across the same issue. Out of curiosity,
>> is this with periodic BC’s in any of the directions? I have had a similar
>> problem (using both 2nd order FD and FE) when using Belos+ML (IFPACK IC)
>> on a 3D anisotropic(from the coefficients) poisson problem. Does SGS
>> smoothing work when using Belos+ML?
>>
>
>
> Hi,
>
> I poked around a little more, and with these settings I can solve the
> system with Belos Block CG:
>  ML ->
>   Reuse Fine Level Smoother = 0
>   Base Method Defaults = SA   [default]
>   ML Settings ->
>    default values = SA
>    max levels = 10
>    prec type = MGV
>    increasing or decreasing = increasing
>    aggregation: type = Uncoupled
>    aggregation: damping factor = 1.333
>    eigen-analysis: type = cg
>    eigen-analysis: iterations = 10
>    smoother: sweeps = 2
>    smoother: damping factor = 1
>    smoother: pre or post = both
>    smoother: type = Chebyshev
>    coarse: type = Amesos-KLU
>    coarse: max size = 128
>    coarse: pre or post = post
>    coarse: sweeps = 1
>
> Changing to symmetric GS seems to blow things up (same error as in the
> first mail) although it appears to depend on the number of CPUs. I'm not
> sure if there's an optimal set of parameters for the 3D Poisson problem?
> Right now I a system with 5594877 is taking 15s (71 CG iterations) to solve
> on 64 CPU cores (2.2GHz Xeon, 8 cores/machine, gigabit ethernet
> interconnect). I'm not sure if it's realistic to expect much faster?
>
> Kind regards,
>
> Bart
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20130509/186008f2/attachment.html 


More information about the Trilinos-Users mailing list