[petsc-users] Problem with SNES convergence

Thu Aug 30 12:55:30 CDT 2018


> On Aug 30, 2018, at 12:49 PM, Fande Kong <fdkong.jd at gmail.com> wrote:
> 
> Hi Barry,
> 
> I haven't had time to look into TS so far. But it is definitely interesting. One simple question would like this: If I have a  simple loop for time steppers, and each step SNES is called. How hard to convert my code to use TS? 

    You would need to look at the examples in ts/examples/tutorials but essentially you need to provide a slightly different version of FormFunction() and FormJacobian() and call TSSolve() etc instead of having your own loop.

    Barry

> 
> Any suggestion? Where should I start from?
> 
> 
> Fande,
> 
> On Thu, Aug 30, 2018 at 11:29 AM Smith, Barry F. <bsmith at mcs.anl.gov> wrote:
> 
>    Note also that the PETSc TS component has a large variety of timesteppers with automatic adaptivity which adjust the time-step for accuracy and convergence. Depending on the exact needs of your time-stepper it might be easier in the long run to use PETSc's time-steppers rather than write your own.
> 
>    Barry
> 
> 
> > On Aug 30, 2018, at 10:46 AM, Ling Zou <Ling.Zou at inl.gov> wrote:
> > 
> > Rahul, please see the logic I used to reduce time step size.
> > Hope this helps you.
> > 
> > Ling
> > 
> > for (timestep = 1; timestep <= N; timestep++) // loop over time steps
> > {
> >   // before trying to solve a time step, 1) it is still not successful; 2) we are not giving up; 3) haven't failed yet
> >   bool give_up = false; bool success = false; bool experienced_fail_this_time_step = false;
> >   // save solutions and old solutions into u_copy and u_old_copy, in case a time step fails and we restart from this saved solutions
> >   VecCopy(u, u_copy); VecCopy(u_old, u_old_copy);
> > 
> >   while ((!give_up) && (!success)) // as long as not successful and not giving up, we solve again with smaller time step
> >   {
> >     if (time_step_size < dt_min) { give_up = true; break; } // ok, bad luck, give up due to time step size smaller than a preset value
> >     if (experienced_fail_this_time_step) { // get the vectors from backups if this is a re-try, i.e., already failed with a larger time step
> >       VecCopy(u_old_copy, u); VecCopy(u_old_copy, u_old);
> >     }
> > 
> >     try {
> >       SNESSolve(snes, NULL, u);
> >       SNESGetConvergedReason(snes, &snes_converged_reason);
> > 
> >       if (snes_converged_reason > 0)  success = true; // yes, snes converged
> >       else { // no, snes did not converge
> >         cutTimeStepSize(); // e.g., dt / 2
> >         experienced_fail_this_time_step = true;
> >       }
> >     }
> >     catch (int err) { // in case your own pieces of code throws an exception
> >       std::cout << "An exception occurred." << std::endl;
> >       success = false;
> >       cutTimeStepSize(); // e.g., dt / 2
> >       experienced_fail_this_time_step = true;
> >     }
> >   }
> > 
> >   if (success) {
> >     // output, print, whatever
> >     // duplicate current solution to old solution in preparing next time step
> >     VecCopy(u, u_old);
> >     // you can increase time step size here, e.g. * 2
> >     increaseTimeStepSize();
> >   }
> > 
> >   if (give_up) {
> >     simulationFailed = true;
> >     std::cerr << "Simulation failed.\n";
> >     //exit(1);// dont exit(1) now, just break the for-loop, let PETSc clean its workspace.
> >     break;
> >   }
> > }
> > 
> > From: Rahul Samala <srahul_05 at yahoo.co.in>
> > Sent: Wednesday, August 29, 2018 10:37:30 PM
> > To: Ling Zou; Smith, Barry F.
> > Cc: PETSc Users List
> > Subject: Re: [petsc-users] Problem with SNES convergence
> >  
> > Thank you Ling, I would definitely like to look at your code for reducing timestep size.
> > Thank you Barry for your inputs.
> > 
> > --
> > Rahul.
> > 
> > On Wednesday, August 29, 2018, 9:02:00 PM GMT+5:30, Smith, Barry F. <bsmith at mcs.anl.gov> wrote:
> > 
> > 
> > 
> > Current time (before start of timestep) 52.5048,    iter=5380     Timestep=864.000000 
> >   0 SNES Function norm 1.650467412595e+05 
> >     0 KSP preconditioned resid norm 3.979123221160e+03 true resid norm 1.650467412595e+05 ||r(i)||/||b|| 1.000000000000e+00
> >     1 KSP preconditioned resid norm 9.178246525982e-11 true resid norm 7.006473307032e-09 ||r(i)||/||b|| 4.245144892632e-14
> >   Linear solve converged due to CONVERGED_RTOL iterations 1
> >   1 SNES Function norm 6.722712947273e+02 
> >   Linear solve did not converge due to DIVERGED_NANORINF iterations 0
> > Nonlinear solve did not converge due to DIVERGED_LINEAR_SOLVE iterations 1
> > 
> >     This usually is an indicator that the LU (ILU) factorization has hit a zero pivot (hence the linear solve has a divide by zero so gives the DIVERGED_NANORINF flag). 
> > 
> >     You can/should call SNESGetConvergedReason() immediately after each SNESSolve(), if the result is negative that means something has failed in the nonlinear solve and you can try cutting the time-step and trying again.
> > 
> >     Good luck,
> > 
> >     Barry
> > 
> > 
> > > On Aug 29, 2018, at 10:11 AM, Ling Zou <Ling.Zou at inl.gov> wrote:
> > > 
> > > 1) My experience is that this kind of bug or sudden death (everything is fine till suddenly something is broken) is very difficult to debug/fix. I looked at your txt files and could not give any quick comments. Maybe PETSc developers have better idea on this.
> > > 2) I do have successful experience on reducing time step size when PETSc fails solving or my own piece of code throws an exception. If you are interested, I can share them.
> > > 
> > > -Ling
> > > From: petsc-users <petsc-users-bounces at mcs.anl.gov> on behalf of Rahul Samala <srahul_05 at yahoo.co.in>
> > > Sent: Wednesday, August 29, 2018 8:36:58 AM
> > > To: PETSc Users List
> > > Subject: [petsc-users] Problem with SNES convergence
> > >  
> > > Hello PetSc users,
> > > 
> > > 1)  I have problem with SNES convergence. I call SNESSolve  in a time loop and use the inbuilt Jacobian feature. The code works fine for about 5380 time steps after which it breaks down. The solution till that point looks fine. I have used newtonls of type l2. (newtontr and others aren't working). Since I have used inbuilt Jacobian feature and the code worked for about 5000 time steps I don't understand the reason for failure, is it an incorrect function evaluation?  Attached are the outputs with -pc_type lu and ilu along with -snes_linesearch_type l2 -snes_converged_reason -snes_monitor -snes_view -ksp_converged_reason -ksp_monitor_true_residual
> > > 
> > > 2)  How to get hold of failure signal, like Nonlinear solve DIVERGED_MAX_IT or DIVERGED_LINEAR_SOLVE so that whenever it occurs I can use a reduced time step and see if the code converges.
> > > 
> > > Thank you,
> > > Rahul.
> > > 
> > > output_ilu.txt
> > > 
> > >     
> > > output_ilu.txt
> > > 
> > > 
> > > 
> > > output_lu.txt
> > > 
> > >     
> > > output_lu.txt
>