> There are some nice options in there to plot the linear solver convergence > and nonlinear solution update - it can be informative.
Thanks for these! I did not realize they were there. (PETSc is so useful when you know what you're looking for!) Using these I was able to find my problem. This generalized alpha is a predictor-corrector method. At the predictor stage I had been updating "local" (ghosted solution values) without updating the normal solution, thus when I entered the newton update and was using the normal solution, it clearly was not doing what it needed to. Thank you both, Jed and Ben, for you thoughts! On to the next bug! =) Andrea > > -Ben > > > ----- Original Message ----- > From: Andrea Hawkins <[email protected]> > To: Kirk, Benjamin (JSC-EG311); libmesh-users > <[email protected]> > Sent: Wed Feb 24 10:18:46 2010 > Subject: Re: [Libmesh-users] Time Stepping > > That wouldn't explain it working after restarts though, would it? > Because with a restart, I'm seeding the initial condition with the > previous solution. So, the linear solver should be using this as an > initial guess, correct? > > Thanks, > Andrea > > > > On Wed, Feb 24, 2010 at 10:14 AM, Kirk, Benjamin (JSC-EG311) > <[email protected]> wrote: >> Since you are rolling your own newton you might need to set the linear >> solver initial guess to 0 (the newton delta) - by default the linear >> implicit system uses the current solution as the krylov initial guess. >> >> Obviously this could be horribly wrong as you converge nonlinearly... >> >> -Ben >> >> >> >> ----- Original Message ----- >> From: Andrea Hawkins <[email protected]> >> To: libmesh-users <[email protected]> >> Sent: Wed Feb 24 09:59:14 2010 >> Subject: [Libmesh-users] Time Stepping >> >> Hello- >> >> I am running a nonlinear time-dependent simulation and am currently >> attempting to implement the generalized alpha time stepping algorithm. >> As default I use PETSc for my linear solves and updates, but this >> method requires I write my own newton update. >> >> I had been setting those options outside of the time loop, but found >> out that the PetscNonlinearSolver was getting cleared after every time >> step so moving that into the time loop solved that problem. However, >> I'm now seeing some very peculiar behavior... The first time step >> things go fine. The second time step, things also seem to go fine, >> however the newton solver actually ends up converging back to the same >> solution it started with so that the third time step starts with the >> same values as the second. I have actually gone into the newton solver >> and seen that the first newton steps leave the initial values but >> converge back. >> >> While this issue could be many things, I have double checked (at >> length) with people familiar with this algorithm that I'm at least >> attempting to implement the correct thing. >> >> Is there something else I should be concerned about getting reset >> every time step? Because when I run it for only one time step and then >> restart and repeat, I do not see this behavior. >> >> Thanks! >> Andrea >> >> ------------------------------------------------------------------------------ >> Download Intel® Parallel Studio Eval >> Try the new software tools for yourself. Speed compiling, find bugs >> proactively, and fine-tune applications for parallel performance. >> See why Intel Parallel Studio got high marks during beta. >> http://p.sf.net/sfu/intel-sw-dev >> _______________________________________________ >> Libmesh-users mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/libmesh-users >> > ------------------------------------------------------------------------------ Download Intel® Parallel Studio Eval Try the new software tools for yourself. Speed compiling, find bugs proactively, and fine-tune applications for parallel performance. See why Intel Parallel Studio got high marks during beta. http://p.sf.net/sfu/intel-sw-dev _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
