[petsc-users] Running AMG libraries as standalone solvers

[Barry Smith]

> On Jul 5, 2023, at 1:18 PM, Khurana, Parv <p.khurana22 at imperial.ac.uk> wrote:
> 
> Hi Barry, 
>  
> Using -ksp_type preonly,and tweaking -pc_hypre_boomeramg_max_iter <it> and -pc_hypre_boomeramg_tol <tol> according to my needs seems to have done the trick for me. 

  Sure, but there is no particular advantage (or disadvantage that I know of) in doing it this way. It is just a question of whether PETSc (via -ksp_type richardson) or hypre (via -pc_hypre_boomeramg_max_iter ) controls the simple iteration (what we call Richardson in PETSc) that wraps around the V-cycle.  I would be surprised if the solve time between -ksp_typre richardson and -pc_hypre_boomeramg_max_iter <it> and -pc_hypre_boomeramg_tol <tol> with similar tolerances was different significantly.

>  
> Credit: https://petsc.org/release/faq/#i-want-to-use-hypre-boomeramg-without-gmres-but-when-i-run-pc-type-hypre-pc-hypre-type-boomeramg-ksp-type-preonly-i-don-t-get-a-very-accurate-answer
>  
> Best
> Parv
>  
> From: petsc-users <petsc-users-bounces at mcs.anl.gov <mailto:petsc-users-bounces at mcs.anl.gov>> On Behalf Of Khurana, Parv
> Sent: 05 July 2023 17:09
> To: Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>>
> Cc: petsc-users at mcs.anl.gov <mailto:petsc-users at mcs.anl.gov>
> Subject: Re: [petsc-users] Running AMG libraries as standalone solvers
>  
> Thank you Mark, Barry for the advice.
>  
> -ksp_type richardson may be what I am looking for. I wanted a simple iterative scheme that allows AMG to be the main work horse – this allows that.
>  
> On using -ksp_type preonly, I observed only 1 KSP iteration and it does not drop the residual norm to my desired level (-ksp_rtol 1e-7). Whereas when I use gmres or Richardson I see 6-10 ksp iterations and the desired residual drop. Is it possible to increase (and fix) the number of KSP iterations with preonly? You mentioned below that it gives only one cycle so I presume it is not possible.
>  
> Geometric mg is not a feasible option for my application as I will be ultimately dealing with 3D, unstructured meshes for hp-finite element methods. Some clever tweaks may allow it (like having sub-structured meshes) but not as of now.
>  
> Best
> Parv
>  
>  
>  
> From: Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> 
> Sent: 05 July 2023 16:51
> To: Khurana, Parv <p.khurana22 at imperial.ac.uk <mailto:p.khurana22 at imperial.ac.uk>>
> Cc: petsc-users at mcs.anl.gov <mailto:petsc-users at mcs.anl.gov>
> Subject: Re: [petsc-users] Running AMG libraries as standalone solvers
>  
> This email from bsmith at petsc.dev <mailto:bsmith at petsc.dev> originates from outside Imperial. Do not click on links and attachments unless you recognise the sender. If you trust the sender, add them to your safe senders list <https://spam.ic.ac.uk/SpamConsole/Senders.aspx> to disable email stamping for this address.
>  
> 
>  
>    BTW: in my experience
>  
> Geometric multigrid generally works great as a standalone solver, that is accelerating it with a Krylov method is not needed, and while it may improve the convergence rate slightly, it may end up taking a bit more time then not using the Krylov solver (due to the extra Krylov solver overhead).
>  
> Algebraic multigrid is usually run with a Krylov solver and running without a Krylov solver, as far as I am aware, generally performs poorly.
>  
>  
> 
> On Jul 5, 2023, at 11:46 AM, Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> wrote:
>  
>  
>   -ksp_type richardson
>  
>   If you use -ksp_typre preonly you effectively get one V-cycle (or whatever cycle type you may have selected) which in general, will give you a more or less garbage answer
>  
>  
>  
> 
> On Jul 5, 2023, at 11:28 AM, Khurana, Parv <p.khurana22 at imperial.ac.uk <mailto:p.khurana22 at imperial.ac.uk>> wrote:
>  
> Hello PETSc users,
>  
> I am fairly inexperienced with PETSc as of now and am new to the mailing list! Thanks for running this channel.
>  
> I seek basic help regarding running AMG routines (BoomerAMG/ML/GAMG). I am trying to compare the performance of solving a Poisson problem using a AMG Preconditioned GMRES iterative solve vs using AMG as the solver. I use PETSc options using the options database keys as of now, and it is connected to a flow solver (Nektar++) I use for my research.
>  
> I currently run the AMG Preconditioned GMRES iterative solve by setting -ksp_type gmres and then specifying the preconditioner I want using, for e.g, -pc_type hypre -pc_hypre_type boomeramg. If I want to use the AMG routine, I am currently setting -ksp_type preonly and the same -pc_type. However, I am not sure if this is correct way to go about it due to two reasons:  a) my solution using AMG as a solver with these options has a larger error than AMG Preconditioned GMRES (which could still be acceptable), and b) I could not find any clear documentation regarding how to use AMG directly as a solver. I saw some hints in the examples herehttps://petsc.org/main/tutorials/handson/# <https://petsc.org/main/tutorials/handson/>, but it hasn’t helped me.
>  
> Any hints on how to use AMG directly as a solver?
>  
> Best
> Parv

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20230705/fb3190e5/attachment-0001.html>