[petsc-users] Reuse KSP

[Smith, Barry F.]

> On Apr 17, 2018, at 4:25 PM, Young, Matthew, Adam <may at bu.edu> wrote:
> 
> That all makes sense. Thanks.
> 
> Is it okay to MatCreate() and MatDestroy() in each time step or am I just defeating the purpose of preserving the KSP?

   You are defeating the purpose. 

   Barry

> 
> Is there a way to tell if the KSP is reusing the factorization (e.g., with -ksp_view)?
> From: Smith, Barry F. <bsmith at mcs.anl.gov>
> Sent: Tuesday, April 17, 2018 5:13:10 PM
> To: Young, Matthew, Adam
> Cc: Matthew Knepley; petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] Reuse KSP
>  
> 
>    If the matrix has a null space then, in general, direct LU factorizations should fail. If they don't fail they can produce an incorrect answer 
> 
>    If your matrix has a null space then you should not remove the  [Set nullspace of A with MatSetNullSpace] 
> 
>    Barry
> 
> > On Apr 17, 2018, at 7:48 AM, Young, Matthew, Adam <may at bu.edu> wrote:
> > 
> > 
> > 
> > 
> > From: Matthew Knepley <knepley at gmail.com>
> > Sent: Tuesday, April 17, 2018 8:38 AM
> > To: Young, Matthew, Adam
> > Cc: petsc-users at mcs.anl.gov
> > Subject: Re: [petsc-users] Reuse KSP
> >  
> > On Tue, Apr 17, 2018 at 8:33 AM, Young, Matthew, Adam <may at bu.edu> wrote:
> > My code (C/C++) uses PETSc to solve a linear system at each iteration in a time-stepping scheme. The PETSc routines are separated out because the scheme that uses them is optional in the larger code. 
> > 
> > At each time step, the PETSc scheme calls a routine that includes:
> > 
> > {
> >   Mat A;
> >   Vec x,b;
> >   KSP ksp;
> > 
> >   [Set up A,x,b]
> > 
> >   KSPCreate(PETSC_COMM_WORLD,&ksp);
> > 
> >   [Build A and b from simulation data]
> > 
> >   [Set nullspace of A with MatSetNullSpace]
> > 
> >   KSPSetOperators(ksp,A,A);
> >   KSPSetFromOptions(ksp);
> >   KSPSolve(ksp,b,x);
> > 
> >   [Write x into an instance of our array class, perform checks, etc.]
> > 
> >   KSPDestroy(&ksp);
> > 
> >   return(0);
> > }
> > 
> > My question is: If I want to use superLU_DIST with the option -mat_superlu_dist_fact SamePattern (or SamePattern_SameRowPerm), should I move KSPCreate/Destroy outside of the time-step loop in the larger code?
> > 
> > 1) Yes
> > 
> > 2) How are you using LU when A has a nullspace?
> > My bad - that code is in there from when I was trying iterative solvers. Should I remove it if I use LU or will PETSc just ignore it?
> > 
> > 3) You can also move out SetFromOptions().
> > Thanks for the advice. So, that means I'll build A and b from simulation data, SetOperators(), Solve(), and carry on?
> > 
> >   Matt
> >  
> > Thanks.
> > ----------------------------
> > Matthew Young
> > PhD Candidate
> > Astronomy Department
> > Boston University
> > 
> > 
> > 
> > 
> > 
> > -- 
> > 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/