[petsc-users] Using SNESSHELL as a wrapper for a CFD solver.
Kenneth C Hall
kenneth.c.hall at duke.edu
Wed May 31 17:42:39 CDT 2023
Matt,
Thanks for this.
Mach number preconditioning is as follows. The Euler or Navier-Stokes equations are written as:
du/dt + M.(dF(u)/dx + dG(u)/dx) = 0
The matrix M is a preconditioning matrix which changes the wave speeds of the convection and the pressure characteristic speeds so that they are close to one another. The equations are no longer time accurate, but converge to a steady states faster. This is especially useful for very low speed flows where the (unpreconditioned) waves travel at very high speeds compared to the convection waves.
Kenneth
From: Matthew Knepley <knepley at gmail.com>
Date: Wednesday, May 31, 2023 at 6:21 PM
To: Barry Smith <bsmith at petsc.dev>
Cc: Kenneth C Hall <kenneth.c.hall at duke.edu>, petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
Subject: Re: [petsc-users] Using SNESSHELL as a wrapper for a CFD solver.
On Wed, May 31, 2023 at 5:51 PM Barry Smith <bsmith at petsc.dev<mailto:bsmith at petsc.dev>> wrote:
Sorry, I wrote to quickly in my last email. You will need to create a SNESSHELL its solve simply calls your solver (for its one iteration) SNESNRICHARSON handles the rest.
Yes, this is a great suggestion. It should work with your current SNESSHELL.
I would also point out that we have the SNESMS, which is the multistage solver from Jameson. You can
use this as a residual smooth with FAS to do something similar to what you have, although I do not know
what Mach number preconditioning is (Jed probably knows).
Thanks,
Matt
On May 31, 2023, at 4:16 PM, Kenneth C Hall <kenneth.c.hall at duke.edu<mailto:kenneth.c.hall at duke.edu>> wrote:
Matt,
Thanks for your quick reply. I think what you say makes sense.
You asked what my code does. The MySolver program performs one iteration of a CFD iteration. The CFD scheme is an explicit scheme that uses multigrid, Mach number preconditioning, and residual smoothing. Typically, I have to call MySolver on the order of 40 to 100 times to get acceptable convergence.
And in fact, I have another version of this Petsc code that uses SNESNGMRES to solve the problem with MySolver providing the residuals as R = N(x) - x. But I would like a version where I am using just MySolver, without any other operations applied to it. So I am trying to plug MySolver into the PETSc system to provide monitoring and other features, and for consistency and comparison to these other (more appropriate!) uses of PETSc.
Thanks.
Kenneth
From: Matthew Knepley <knepley at gmail.com<mailto:knepley at gmail.com>>
Date: Wednesday, May 31, 2023 at 3:48 PM
To: Kenneth C Hall <kenneth.c.hall at duke.edu<mailto:kenneth.c.hall at duke.edu>>
Cc: petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov> <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>>
Subject: Re: [petsc-users] Using SNESSHELL as a wrapper for a CFD solver.
On Wed, May 31, 2023 at 3:21 PM Kenneth C Hall <kenneth.c.hall at duke.edu<mailto:kenneth.c.hall at duke.edu>> wrote:
Hi,
I am doing a number of problems using PETSc/SLEPc, but I also work on some non-PETSc/SLEPc flow solvers. I would like to use PETSc as a wrapper for this non-PETSc flow solver for compatibility, so I can use the tolerance monitoring, options, viewers, and for direct comparison to PETSc methods I am using.
Here is what I am trying to do… I have a CFD solver that iterates with a nonlinear iterator of the form x := N(x). This can be expressed in a fortran routine of the form,
SUBROUTINE MySolver(x)
or
SUBROUTINE MySolver(x,y)
In the first case, x is over written. In the second, y = N(x). In any event, I want to do something like what is shown in the subroutine at the bottom of this email.
The code below “works” in the sense that MySolver is called, but it is called exactly *once*. But MyMonitor and MyConverged are *not* called. Again, I want to iterate so MySolver should be called many times, as should MyMonitor and MyConverged.
The SNESolve() method is called once per nonlinear solve, just as KSPSolve() is called once per linear solve. There may be iteration inside the method, but that is handled inside the particular implementation. For example, both Newton's method and Nonlinear Conjugate Gradient iterate, but the iteration is internal to both, and they both call the monitor and convergence test at each internal iterate.
So, if your nonlinear solver should iterate, it should happen inside the SNESSolve call for the SNESSHELL object. Does this make sense? What does your solver do?
Thanks,
Matt
The SNESView before and after SNESSolve looks like this:
SNES Object: 1 MPI process
type: shell
SNES has not been set up so information may be incomplete
maximum iterations=50, maximum function evaluations=10000
tolerances: relative=1e-50, absolute=1e-10, solution=1e+06
total number of function evaluations=0
norm schedule ALWAYS
SNES Object: 1 MPI process
type: shell
maximum iterations=50, maximum function evaluations=10000
tolerances: relative=1e-50, absolute=1e-10, solution=1e+06
total number of function evaluations=0
norm schedule ALWAYS
Any suggestions on how to do what I am trying to accomplish?
Thanks.
Kenneth Hall
#include <petsc/finclude/petsc.h>
#include "macros.h"
MODULE SolveWithSNESShell_module
USE MyPetscModule
CONTAINS
!
!====================================================================================================
SUBROUTINE MySolver(snes, x, ierr)
!====================================================================================================
!!
!!
!====================================================================================================
!
USE MyPetscModule
IMPLICIT NONE
!
!.... declared passed variables
SNES :: snes
Vec :: x
PetscErrorCode :: ierr
!
!.... code to find residual x := N(x)
!.... (or alternatively y := N(x))
END SUBROUTINE MySolver
!
!====================================================================================================
SUBROUTINE MyMonitor(snes, its, rnorm, ierr)
!====================================================================================================
!!
!!
!====================================================================================================
!
USE MyPetscModule
IMPLICIT NONE
!
!.... Declare passed variables
SNES :: snes
PetscInt :: its
PetscReal :: rnorm
PetscErrorCode :: ierr
!
!.... Code to print out convergence history
!.... Code to print out convergence history
END SUBROUTINE MyMonitor
!====================================================================================================
SUBROUTINE MyConverged(snes, it, xnorm, ynorm, znorm, reason, ierr)
USE MyPetscModule
IMPLICIT NONE
SNES :: snes
PetscInt :: it,ctx
PetscReal :: xnorm, ynorm, znorm
KSPConvergedReason :: reason
PetscErrorCode :: ierr
! ... add convergence test here ...
! set reason to a positive value if convergence has been achieved
END SUBROUTINE MyConverged
END MODULE SolveWithSNESShell_module
!
!====================================================================================================
SUBROUTINE SolveWithSNESShell
!====================================================================================================
!!
!!
!====================================================================================================
!
USE SolveWithSNESShell_module
IMPLICIT NONE
!
!.... Declare passed variables
INTEGER :: level_tmp
!
!.... Declare local variables
INTEGER :: iz
INTEGER :: imax
INTEGER :: jmax
INTEGER :: kmax
SNES :: snes
KSP :: ksp
Vec :: x
Vec :: y
PetscViewer :: viewer
PetscErrorCode :: ierr
PetscReal :: rtol = 1.0D-10 !! relative tolerance
PetscReal :: atol = 1.0D-50 !! absolute tolerance
PetscReal :: dtol = 1.0D+06 !! divergence tolerance
PetscInt :: maxits = 50
PetscInt :: maxf = 10000
character(len=1000):: args
!
!.... count the number of degrees of freedom.
level = level_tmp
n = 0
DO iz = 1, hb(level)%nzone
imax = hb(level)%zone(iz)%imax - 1
jmax = hb(level)%zone(iz)%jmax - 1
kmax = hb(level)%zone(iz)%kmax - 1
n = n + imax * jmax * kmax
END DO
n = n * neqn
!
!.... Initialize PETSc
PetscCall(PetscInitialize(PETSC_NULL_CHARACTER, ierr))
!
!.... Log
PetscCall(PetscLogDefaultBegin(ierr))
!
!.... Hard-wired options.
! PetscCall(PetscOptionsInsertString(PETSC_NULL_OPTIONS, "command line style option here" , ierr))
!
!.... Command line options.
call GET_COMMAND(args)
PetscCall(PetscOptionsInsertString(PETSC_NULL_OPTIONS, args, ierr))
!
!.... view command line table
PetscCall(PetscViewerASCIIOpen(PETSC_COMM_SELF, PETSC_VIEWER_STDOUT_SELF, viewer, ierr))
PetscCall(PetscOptionsView(PETSC_NULL_OPTIONS, viewer, ierr))
PetscCall(PetscViewerDestroy(viewer, ierr))
!
!.... Create PETSc vectors
PetscCall(VecCreateSeq(PETSC_COMM_SELF, n, x, ierr))
PetscCall(VecCreateSeq(PETSC_COMM_SELF, n, y, ierr))
PetscCall(VecSet(x, 0.0d0, ierr))
PetscCall(VecSet(y, 0.0d0, ierr))
!.... SNES context
PetscCall(SNESCreate(PETSC_COMM_SELF, snes, ierr))
PetscCall(SNESSetType(snes, SNESSHELL, ierr))
PetscCall(SNESShellSetSolve(snes, MySolver, ierr))
!!!! PetscCall(SNESSetFunction(snes, x, MySolver, PETSC_NULL_INTEGER, ierr))
!!!! this line causes a segmentation error if uncommented.
PetscCall(SNESSetConvergenceTest(snes, MyConverged, 0, PETSC_NULL_FUNCTION, ierr))
!
!.... Set SNES options
PetscCall(SNESSetFromOptions(snes, ierr))
!.... Set tolerances
PetscCall(SNESSetTolerances(snes, rtol, atol, dtol, maxits, maxf, ierr))
!
!.... SNES montior
PetscCall(SNESMonitorSet(snes, MyMonitor, PETSC_NULL_INTEGER, PETSC_NULL_FUNCTION,ierr))
!
!.... Set the initial solution
CALL HBToVecX(x)
!
!.... View snes context
PetscCall(SNESView(snes, viewer, ierr))
!
!.... Solve SNES problem
PetscCall(SNESSolve(snes, PETSC_NULL_VEC, x, ierr))
!
!.... View snes context
PetscCall(SNESView(snes, viewer, ierr))
!
!.... dump the logs
! call PetscLogDump(ierr) ! Why does this cause error
!
!.... Destroy PETSc objects
PetscCall(SNESDestroy(snes, ierr))
PetscCall(VecDestroy(x, ierr))
PetscCall(VecDestroy(y, ierr))
PetscCall(PetscViewerDestroy(viewer, ierr))
!
!.... Finish
PetscCall(PetscFinalize(ierr))
END SUBROUTINE SolveWithSNESShell
--
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/<https://urldefense.com/v3/__http:/www.cse.buffalo.edu/*knepley/__;fg!!OToaGQ!sE2W1qI_dqcEHL1dOCnJ3Rdv9TATFVDBiBqx_tlQsOnjvGF7StDjsmVcm9Qkfe4XcTFOBtVjtFl5om07Rdjw$>
--
What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/<https://urldefense.com/v3/__http:/www.cse.buffalo.edu/*knepley/__;fg!!OToaGQ!s-GnaelCE7pDC_xWpNI3mM5IDxDYowWhn_fSvgL-9JPgdORySPUEHjp-rhUcrlLnhv8LCx38Hr_TCa18TRGy$>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20230531/3cc04ed1/attachment-0001.html>
More information about the petsc-users
mailing list