Re: [petsc-users] Question about preconditioner

[Barry Smith]

   If your  matrix has the form 

(   A   B )
(   C   0 )

then often PCFIELDSPLIT can be a useful preconditioner with the option 
-pc_fieldsplit_detect_saddle_point



> On Feb 16, 2023, at 2:42 AM, user_gong Kim  wrote:
> 
>  
> Hello,
> 
>  
> There are some questions about some preconditioners.
> 
> The questions are from problem Au=b. The global matrix A has zero value 
> diagonal terms.
> 
> 1. Which preconditioner is preferred for matrix A which has zero value in 
> diagonal terms?
> The most frequently used basic 2 preconditioners are jacobi and SOR (gauss 
> seidel). As people knows both methods should have non zero diagonal terms. 
> Although the improved method is applied in PETSc, jacobi can also solve the 
> case with zero diagonal term, but I ask because I know that it is not 
> recommended.
> 
> 2. Second question is about running code with the two command options 
> below in a single process.
> 1st command : -ksp_type gmres -pc_type bjacobi -sub_pc_type jacobi
> 2nd command : -ksp_type gmres -pc_type hpddm -sub_pc_type jacobi
> When domain decomposition methods such as bjacobi or hpddm are parallel, the 
> global matrix is divided for each process. As far as I know, running it in a 
> single process should eventually produce the same result if the sub pc type 
> is the same. However, in the second option, ksp did not converge.
> In this case, I wonder how to analyze the situation.
> How can I monitor and see the difference between the two?
> 
>  
>  
> Thanks,
> 
> Hyung Kim
> 

Re: [petsc-users] Question about preconditioner

[Matthew Knepley]

On Thu, Feb 16, 2023 at 2:43 AM user_gong Kim  wrote:

>
>
> Hello,
>
>
>
> There are some questions about some preconditioners.
>
> The questions are from problem Au=b. The global matrix A has zero value
> diagonal terms.
>
> 1. Which preconditioner is preferred for matrix A which has zero
> value in diagonal terms?
> The most frequently used basic 2 preconditioners are jacobi and SOR (gauss
> seidel). As people knows both methods should have non zero diagonal terms.
> Although the improved method is applied in PETSc, jacobi can also solve the
> case with zero diagonal term, but I ask because I know that it is not
> recommended.
>
> 2. Second question is about running code with the two command options
> below in a single process.
> 1st command : -ksp_type gmres -pc_type bjacobi -sub_pc_type jacobi
> 2nd command : -ksp_type gmres -pc_type hpddm -sub_pc_type jacobi
> When domain decomposition methods such as bjacobi or hpddm are parallel,
> the global matrix is divided for each process. As far as I know, running it
> in a single process should eventually produce the same result if the sub pc
> type is the same. However, in the second option, ksp did not converge.
> In this case, I wonder how to analyze the situation.
> How can I monitor and see the difference between the two?
>
> Pierre is correct. I will just note that the best way to use PETSc is
generally to find the preconditioner you need
by looking in literature for what has been successful for other people, and
then reproducing it in PETSc, which
should be easy.

  Thanks,

 Matt


>
>
> Thanks,
>
> Hyung Kim
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ 

Re: [petsc-users] Question about preconditioner

[Pierre Jolivet]

> On 16 Feb 2023, at 8:43 AM, user_gong Kim  wrote:
> 
> 
>  
> 
> Hello,
> 
>  
> 
> There are some questions about some preconditioners.
> 
> The questions are from problem Au=b. The global matrix A has zero value 
> diagonal terms.
> 
> 1. Which preconditioner is preferred for matrix A which has zero value in 
> diagonal terms?
> 

This question has not a single answer. It all depends on where your A and b are 
coming from.

> The most frequently used basic 2 preconditioners are jacobi and SOR (gauss 
> seidel).
> 
They are not the most frequently used. And rightfully so, as they very often 
can’t handle non-trivial systems.

> As people knows both methods should have non zero diagonal terms. Although 
> the improved method is applied in PETSc, jacobi can also solve the case with 
> zero diagonal term, but I ask because I know that it is not recommended.
> 
> 2. Second question is about running code with the two command options 
> below in a single process.
> 1st command : -ksp_type gmres -pc_type bjacobi -sub_pc_type jacobi
> 2nd command : -ksp_type gmres -pc_type hpddm -sub_pc_type jacobi
> When domain decomposition methods such as bjacobi or hpddm are parallel, the 
> global matrix is divided for each process. As far as I know, running it in a 
> single process should eventually produce the same result if the sub pc type 
> is the same. However, in the second option, ksp did not converge.
> 
1st command: it’s pointless to couple PCBJACOBI with PCJABOCI, it’s equivalent 
to only using PCJABOBI.
2nd command: it’s pointless to use PCHPDDM if you don’t specify in some way how 
to coarsen your problem (either algebraically or via an auxiliary operator). 
You just have a single level (equivalent to PCBJACOBI), but its options are 
prefixed by -pc_hpddm_coarse_ instead of -sub_
Again, both sets of options do not make sense.
If you want, you could share your A and b (or tell us what you are 
discretizing) and we will be able to provide a better feedback.

Thanks,
Pierre

> In this case, I wonder how to analyze the situation.
> How can I monitor and see the difference between the two?
> 
>  
> 
>  
> 
> Thanks,
> 
> Hyung Kim

[petsc-users] Question about preconditioner

[user_gong Kim]

Hello,



There are some questions about some preconditioners.

The questions are from problem Au=b. The global matrix A has zero value
diagonal terms.

1. Which preconditioner is preferred for matrix A which has zero value
in diagonal terms?
The most frequently used basic 2 preconditioners are jacobi and SOR (gauss
seidel). As people knows both methods should have non zero diagonal terms.
Although the improved method is applied in PETSc, jacobi can also solve the
case with zero diagonal term, but I ask because I know that it is not
recommended.

2. Second question is about running code with the two command options
below in a single process.
1st command : -ksp_type gmres -pc_type bjacobi -sub_pc_type jacobi
2nd command : -ksp_type gmres -pc_type hpddm -sub_pc_type jacobi
When domain decomposition methods such as bjacobi or hpddm are parallel,
the global matrix is divided for each process. As far as I know, running it
in a single process should eventually produce the same result if the sub pc
type is the same. However, in the second option, ksp did not converge.
In this case, I wonder how to analyze the situation.
How can I monitor and see the difference between the two?





Thanks,

Hyung Kim