[petsc-users] help about my first petsc program

[Jed Brown]

On Wed, 23 Jun 2010 12:45:37 -0400, Xuan YU <xxy113 at psu.edu> wrote:
> I changed dt=0.4 and the Monitor function, so GL doesn't work.

Choosing a smaller dt (only an initial value in this case) fixes that
problem, but I still see some difficulties unless I give
-ts_gl_max_order 2 (seems like I don't have enough stabilization for the
higher-order error estimators).  But feel free to keep using the simpler
methods, I'm sure they will be more predictable.

> PetscErrorCode Monitor(TS ts,PetscInt step,PetscReal time,Vec u, void  
> *ctx)
> {
>    PetscScalar *y;
>    PetscReal dt;
>    TSGetTimeStep(ts,&dt);
>    VecGetArray(u,&y);
>    PetscPrintf(PETSC_COMM_WORLD,"At t =%e  y =%e  %e  %e,  
> \n",time,y[0],y[1],y[2]);
>    VecRestoreArray(u,&y);
>    if(step>0)dt=time*9;
>    TSSetTimeStep(ts,dt);
>    return 0;
> }

TSGL has an adaptive controller, based on a local error estimator.  It
surely won't work right if the monitor changes it (because it doesn't
happen in the correct place; GL can't change the step size arbitrarily,
it needs to know the size of the next step when it "completes" the
current one).  You should be able to TSGLAdaptRegisterDynamic() you own
adaptive scheme, there is currently NONE, SIZE (only step size, not
order), and BOTH (default).  If you do this, and it works better than
the builtin ones, let me know and I'll add it to the library.

> Why Petsc is faster than cvode? How fast it will be?

It might not be faster, but at least you can get back CVODE as a runtime
option.  And since PETSc offers a lot more preconditioning options, you
can probably beat it that way.  Whether PETSc-native integrators
(e.g. TSTHETA, TSGL) are faster or not is very problem dependent.

Jed