November 2011 - petsc-dev - mailman3.cels.anl.gov

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Jed Brown 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 17:37, Matthew Knepley <knepley(a)gmail.com> wrote: > b) FormFunction and FormJacobian in a SNES could be used by an enclosing > TS to do pseudo-timestepping This can't work. What equation would the FormFunction and FormJacobian correspond to? The TS *defines* the equation that needs to be solved. The user provides information according to an interface. The TS interface is more general, and really is required for transient things. The SNESFormFunction isn't enough because you haven't identified positivity and constraints. Add that transient term to all variables equally and you are going to see some confusing stuff like huge conservation error and incorrect boundary conditions, such that you could easily end up on the wrong branch or mis-diagnose a limit cycle versus steady state.

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Matthew Knepley 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 5:13 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > On Nov 15, 2011, at 3:48 PM, Jed Brown wrote: > > > On Tue, Nov 15, 2011 at 15:44, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > How about a SNES level interface that backs into the TS for the > implementation? The thing is "if one is solving a nonlinear problem they > should use the nonlinear solver interface" not a completely different > interface that exists for a different purpose. > > > > Well, I don't choose KSP, SNES, or TS based on what kind of problem I'm > trying to solve, > > I do :-) > > > > I choose it based on how I want to describe the problem. So I might use > SNES for a linear problem because I want to write a function to evaluate > the RHS and a function to evaluate the Jacobian. And I might use TS for a > steady problem because I want a principled way to do damping and to > accelerate certain regions. > > Hmm, if I write my code to use SNES to solve a nonlinear system and > decide I want to instead use Ptc I have to write new code and a whole knew > interface. > > If I wrote my code to use Jacobi and decided I wanted to try ILU should > I have to write a new code? Obviously not so why is your case different? I > just add a mass matrix (for example) and say I want to use Ptc with SNES or > don't even add a mass matrix just say I want to use Ptc. This does expose a problem with our solvers. They should fit together better than they do. I would like to see them telescope so that a) A and b set into a KSP could be used by an enclosing SNES to do iterative refinement b) FormFunction and FormJacobian in a SNES could be used by an enclosing TS to do pseudo-timestepping This way, the user does no more work than create a solver wrapper, and we get interface separation. Matt > > > > > I think it would be more confusing to use SNES because then we would > need to augment the SNES interface so the user can provide the IJacobian. > Having another function signature that the user has to implement is BAD. > > I think using TS to solve a nonlinear equation is more confusing. > > This needs a few use cases to clarify > > Barry > > -- 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

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Jed Brown 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 17:34, Matthew Knepley <knepley(a)gmail.com> wrote: > I am not sure I like this since it creates a circular dependency in the > libraries. Like how MATSCHURCOMPLEMENT creates a circular dependency? Which it doesn't because the implementation lives in KSP. This would be the same, the implementation would live in TS. But I don't like because the user needs to provide something else anyway, and the most natural interface for that thing they need to provide is TSIJacobian.

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Matthew Knepley 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 3:44 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > On Nov 15, 2011, at 3:40 PM, Jed Brown wrote: > > > On Tue, Nov 15, 2011 at 15:36, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > I think it is best if the TS example be cleaned up to do everything the > SNES example did and not have both around. People will stumble upon the > SNES one and use it as a template for their work when they should use the > TS one. > > > > Okay, I'll document the options to reproduce the paper's results. > > > > > > I would like the TS pseudo time-stepping stuff to actually be a SNES > solver. That is SNESType SNESPSEUDOTS or something and not be part of TS. > This is controversial hence nothing has been done to make it this way. Some > people may argue it is the wrong approach. > > > > Well, the way in which the transient term is added is really quite > important. Especially since Ptc often uses some sort of local time > stepping, I don't see a good way to produce that without creating the TS > interface. Since the interface is TS, I think it belongs there. > > How about a SNES level interface that backs into the TS for the > implementation? The thing is "if one is solving a nonlinear problem they > should use the nonlinear solver interface" not a completely different > interface that exists for a different purpose. I am not sure I like this since it creates a circular dependency in the libraries. Matt > > Barry > > > -- 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

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Jed Brown 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 17:13, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > Hmm, if I write my code to use SNES to solve a nonlinear system and > decide I want to instead use Ptc I have to write new code and a whole knew > interface. > > If I wrote my code to use Jacobi and decided I wanted to try ILU should > I have to write a new code? > Well, suppose you were allergic to assembled matrices, so you had a MatShell and implemented Jacobi so that you could use Chebychev/Jacobi smoothers. Now you want to do ILU. Sounds like you have to either assemble a matrix or implement MatFactor*. (Coloring is a special trick, so let's pretend it doesn't exist for the moment.) > Obviously not so why is your case different? I just add a mass matrix (for > example) and say I want to use Ptc with SNES or don't even add a mass > matrix just say I want to use Ptc. > The problem is that it's often not physical for the mass matrix to be the identity. The lid-driven cavity is a classical example of this, the transient term should only be added to the vorticity and temperature equations. It's physically nonsense to add a transient term to the velocity (in velocity-vorticity form). So sure, we can "guess" a bunch of stuff that you didn't tell us to automatically put together a damped Newton method, but to do Ptc, you have to provide the pseudo-transient form. It would be far more silly to tell someone that to move from goofy damped Newton to Ptc with a proper transient term involves rewriting for a different interface. But if you don't do that, then you are going to have to add a SNESSetPseudotransientJacobian() or some such special-purpose interface. Use TS if your problem has a meaningful transient form, it's the more general interface. > > I think it would be more confusing to use SNES because then we would > need to augment the SNES interface so the user can provide the IJacobian. > Having another function signature that the user has to implement is BAD. > > I think using TS to solve a nonlinear equation is more confusing. > If you add the interface, then YOU are going to be responsible for explaining to users why there are two ways to add a transient term, but only one can be used for time stepping.

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Barry Smith 15 Nov '11

15 Nov '11

On Nov 15, 2011, at 3:48 PM, Jed Brown wrote: > On Tue, Nov 15, 2011 at 15:44, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > How about a SNES level interface that backs into the TS for the implementation? The thing is "if one is solving a nonlinear problem they should use the nonlinear solver interface" not a completely different interface that exists for a different purpose. > > Well, I don't choose KSP, SNES, or TS based on what kind of problem I'm trying to solve, I do :-) > I choose it based on how I want to describe the problem. So I might use SNES for a linear problem because I want to write a function to evaluate the RHS and a function to evaluate the Jacobian. And I might use TS for a steady problem because I want a principled way to do damping and to accelerate certain regions. Hmm, if I write my code to use SNES to solve a nonlinear system and decide I want to instead use Ptc I have to write new code and a whole knew interface. If I wrote my code to use Jacobi and decided I wanted to try ILU should I have to write a new code? Obviously not so why is your case different? I just add a mass matrix (for example) and say I want to use Ptc with SNES or don't even add a mass matrix just say I want to use Ptc. > > I think it would be more confusing to use SNES because then we would need to augment the SNES interface so the user can provide the IJacobian. Having another function signature that the user has to implement is BAD. I think using TS to solve a nonlinear equation is more confusing. This needs a few use cases to clarify Barry

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Jed Brown 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 15:44, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > How about a SNES level interface that backs into the TS for the > implementation? The thing is "if one is solving a nonlinear problem they > should use the nonlinear solver interface" not a completely different > interface that exists for a different purpose. Well, I don't choose KSP, SNES, or TS based on what kind of problem I'm trying to solve, I choose it based on how I want to describe the problem. So I might use SNES for a linear problem because I want to write a function to evaluate the RHS and a function to evaluate the Jacobian. And I might use TS for a steady problem because I want a principled way to do damping and to accelerate certain regions. I think it would be more confusing to use SNES because then we would need to augment the SNES interface so the user can provide the IJacobian. Having another function signature that the user has to implement is BAD.

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Barry Smith 15 Nov '11

15 Nov '11

On Nov 15, 2011, at 3:40 PM, Jed Brown wrote: > On Tue, Nov 15, 2011 at 15:36, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > I think it is best if the TS example be cleaned up to do everything the SNES example did and not have both around. People will stumble upon the SNES one and use it as a template for their work when they should use the TS one. > > Okay, I'll document the options to reproduce the paper's results. > > > I would like the TS pseudo time-stepping stuff to actually be a SNES solver. That is SNESType SNESPSEUDOTS or something and not be part of TS. This is controversial hence nothing has been done to make it this way. Some people may argue it is the wrong approach. > > Well, the way in which the transient term is added is really quite important. Especially since Ptc often uses some sort of local time stepping, I don't see a good way to produce that without creating the TS interface. Since the interface is TS, I think it belongs there. How about a SNES level interface that backs into the TS for the implementation? The thing is "if one is solving a nonlinear problem they should use the nonlinear solver interface" not a completely different interface that exists for a different purpose. Barry

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Jed Brown 15 Nov '11

15 Nov '11

On Tue, Nov 15, 2011 at 15:36, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > I think it is best if the TS example be cleaned up to do everything the > SNES example did and not have both around. People will stumble upon the > SNES one and use it as a template for their work when they should use the > TS one. > Okay, I'll document the options to reproduce the paper's results. > > I would like the TS pseudo time-stepping stuff to actually be a SNES > solver. That is SNESType SNESPSEUDOTS or something and not be part of TS. > This is controversial hence nothing has been done to make it this way. Some > people may argue it is the wrong approach. > Well, the way in which the transient term is added is really quite important. Especially since Ptc often uses some sort of local time stepping, I don't see a good way to produce that without creating the TS interface. Since the interface is TS, I think it belongs there.

1 0

Re: [petsc-dev] snes/examples/tutorials/ex27.c
by Barry Smith 15 Nov '11

15 Nov '11

On Nov 15, 2011, at 3:25 PM, Jed Brown wrote: > http://petsc.cs.iit.edu/petsc/petsc-dev/rev/c3192eb7cd93 > > I would rather not delete this example because it is the example code accompanying a paper that I like. I wrote essentially the same thing in ts/examples/tutorials/ex26.c (which uses TSPseudo), but having the simple, executable example that exactly reproduces the figure in the paper is still useful. Using TS is somewhat less clear because you have to look at library code to see the algorithm, but I could also live with simply documenting how to get the same results with ts/.../ex26.c. > > Preference? I think it is best if the TS example be cleaned up to do everything the SNES example did and not have both around. People will stumble upon the SNES one and use it as a template for their work when they should use the TS one. > > As a general rule, many elliptic problems arise as the steady state of some dynamical system. If we write the code using the TS interface, then we can always solve the steady problem by taking an infinite step size with nonlinearly implicit Euler. Should we recommend that all such problems be written using the TS interface? I would like the TS pseudo time-stepping stuff to actually be a SNES solver. That is SNESType SNESPSEUDOTS or something and not be part of TS. This is controversial hence nothing has been done to make it this way. Some people may argue it is the wrong approach. Barry

1 0