[petsc-users] asymmetric errors at X+/Y+ boundary on periodic box mesh

[Yann Jobic]

Dear all,

Let me be more precise, all the relevant steps that i'm doing are those :
   const char *opts =
     "-dm_plex_box_faces 32,32 -dm_plex_simplex 0 "
     "-dm_plex_box_bd periodic,periodic -dm_distribute_overlap 1 "
     "-dm_plex_adj_cone -dm_plex_adj_closure 0";
   PetscCall(PetscOptionsInsertString(NULL, opts));

   PetscCall(DMCreate(PETSC_COMM_WORLD, &dm));
   PetscCall(DMSetType(dm, DMPLEX));
   PetscCall(DMSetFromOptions(dm));
   {
     DM g;
     PetscCall(DMPlexConstructGhostCells(dm, NULL, NULL, &g));
     PetscCall(DMDestroy(&dm));
     dm = g;
   }

   /* FV: leastsquares + vanleer. */
   PetscCall(PetscFVCreate(PETSC_COMM_WORLD, &fvm));
   PetscCall(PetscFVSetType(fvm, PETSCFVLEASTSQUARES));
   PetscCall(PetscFVSetNumComponents(fvm, 1));
   PetscCall(PetscFVSetSpatialDimension(fvm, 2));
   PetscCall(PetscFVSetComputeGradients(fvm, PETSC_TRUE));
   {
     PetscLimiter lim;
     PetscCall(PetscFVGetLimiter(fvm, &lim));
     PetscCall(PetscLimiterSetType(lim, PETSCLIMITERVANLEER));
   }
   PetscCall(DMAddField(dm, NULL, (PetscObject)fvm));
   PetscCall(DMCreateDS(dm));
   PetscCall(DMGetDS(dm, &ds));
   PetscCall(PetscDSSetRiemannSolver(ds, 0, Riemann));

   /* IC + TS. */
   PetscCall(DMCreateGlobalVector(dm, &U));
   PetscCall(PetscObjectSetName((PetscObject)U, "phi"));
   {
     PetscErrorCode (*f[1])(PetscInt, PetscReal, const PetscReal[],
                            PetscInt, PetscScalar[], void*) = {Sine};
     void *fctx[1] = {NULL};
     PetscCall(DMProjectFunction(dm, 0.0, f, fctx, INSERT_ALL_VALUES, U));
   }

   PetscCall(TSCreate(PETSC_COMM_WORLD, &ts));
   PetscCall(TSSetType(ts, TSSSP));
   PetscCall(TSSSPSetType(ts, TSSSPRKS3));
   PetscCall(TSSetDM(ts, dm));
   PetscCall(DMTSSetBoundaryLocal(dm, DMPlexTSComputeBoundary, NULL));
   PetscCall(DMTSSetRHSFunctionLocal(dm, DMPlexTSComputeRHSFunctionFVM, 
NULL));
   PetscCall(TSSetTimeStep(ts, 1e-3));
   PetscCall(TSSetMaxTime(ts, 0.01));
   PetscCall(TSSetExactFinalTime(ts, TS_EXACTFINALTIME_MATCHSTEP));
   PetscCall(TSSetSolution(ts, U));
   PetscCall(TSSolve(ts, U));

It's the important parts of the minimal working example that i gave in 
the first email.

Maybe i'm doing something wrong for periodic boundaries ?
I may have forgot an important step ?
Maybe there are CI tests that validate the order of accuracy of 
DMPlexTSComputeRHSFunctionFVM on a periodic hex DMPlex with a 
smooth manufactured solution? I may have missed it.


Many thanks,

Yann

Le 22/05/2026 à 15:50, Yann Jobic a écrit :
> Ce mail provient de l'extérieur, restons vigilants
> 
> Dear all,
> 
> I'm writing a small finite-volume scalar advection code on a 2D periodic 
> DMPlex hex mesh, using the same pipeline as src/ts/tutorials/ex11.c 
> (PetscFV + PetscLimiter + DMPlexTSComputeRHSFunctionFVM + TSSSP).
> 
> After only 10 time steps from a smooth periodic initial condition 
> phi_0(x,y) = sin(2 pi x) * sin(2 pi y), with constant velocity a = 
> (1,1), on a 32x32 hex periodic box, the error |phi - phi_exact| is 
> strongly asymmetric: it is roughly 0.18 in the cells touching x=1 and 
> y=1, but about a tenth of that everywhere else (including near x=0 and 
> y=0). The exact solution is just phi_0 translated by a*t and stays smooth.
> 
> I'm using PETSc 3.24.4 (release) and 3.25.1 (release).
> 
> I'm attaching a 170-line MWE (mwe.c, no command-line options needed) and 
> a screenshot of the problem. The MWE prints
> t=1.000e-02  Linf=1.7897e-01
> 
> Is there an additional setup step that I'm missing for the periodic hex 
> case (e.g. an explicit call about face localization, or about the SF 
> after DMPlexConstructGhostCells)?
> 
> Thanks,
> 
> Yann