[petsc-users] How to use multigrid?

[Mark F. Adams]

If your machine can "scale" a dot product then it can scale AMG.  That said,  implementing AMG to scale well takes work.  GAMG should be able to go to 10^4 processes.  We are working on some scaling problems in the setup code and I'm not too sure what and where GAMG will break as you scale up.  I'd be curious to know.  hypre should scale pretty well.


On Nov 9, 2012, at 12:43 PM, Thomas Witkowski <Thomas.Witkowski at tu-dresden.de> wrote:

> :) It was just a question whether is makes sense to run my solver in this range of number of cores, or whether AMG by itself will not scale here anymore. As my computational time is limited, I have to think carefully when using 10^4 cores.
> 
> Thomas
> 
> 
> Am 09.11.2012 18:40, schrieb Jed Brown:
>> Don't speculate, send -log_summary and information about the problem and machine. Some setup operations degrade, but solves should scale pretty well.
>> 
>> On Fri, Nov 9, 2012 at 11:37 AM, Thomas Witkowski <thomas.witkowski at tu-dresden.de> wrote:
>> While we're at it, can some of you tell me some world about parallel scaling of algebraic multigrid methods? In most of my codes I use them to precondition some simple blocks, e.g. Laplace matrix. I'm pretty sure that parallel scaling of my solver is limited mostly by scaling of the AMG method which is used. What are the expectations, when going to 10^3 or 10^4 cores?
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20121109/07adef26/attachment.html>