Fixed an issue where applying DirectSolver to StateIndependentsComp was breaking when used with other linear solvers under MPI. #1020
Conversation
… transcriptions. Phase now has default_linear_solver and default_nonlinear_solver options like trajectory.
… other solvers potentially unknown to us
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
…e of NonlinearBlockJac/PetscKrylov. The solvers are needed in the presence of MPI regardless of number of procs.
…kJac if there are direct connections.
Added auto_solvers options to Trajectory and Phases, allowing users to disable the behavior if desired. Added parallel_phases option to Trajectory, which toggles whether the top level phase container is a parallel group.
| @@ -684,7 +684,7 @@ def configure_solvers(self, phase): | |||
| phase : dymos.Phase | |||
| The phase object to which this transcription instance applies. | |||
There was a problem hiding this comment.
just noting the docstring here doesn't include the 'requires_solvers' arg
dymos/trajectory/trajectory.py
Outdated
|
|
||
| warn = False | ||
| if self._has_connected_phases and isinstance(self.phases, om.ParallelGroup) and MPI and self.comm.size > 1: |
There was a problem hiding this comment.
Since self.comm.size will never be > 1 if MPI is not active, you don't really need the and MPI in the if check.
| if requires_solvers is not None: | ||
| req_solvers.update(requires_solvers) | ||
|
|
||
| reasons_txt = '\n'.join(f' {key}: {val}' for key, val in req_solvers.items()) |
There was a problem hiding this comment.
This could move inside the if any(req_solvers.values()) block
| if requires_solvers is not None: | ||
| req_solvers.update(requires_solvers) | ||
|
|
||
| reasons_txt = '\n'.join(f' {key}: {val}' for key, val in req_solvers.items()) |
There was a problem hiding this comment.
Intersting that the user can put whatever they want in this sentence. Looks like it also prints out for every phase in the model?
There was a problem hiding this comment.
The intention was to try to include a "reason" in the warning that gets printed out, so allowing the string in here seemed like the quickest way to do it.
And yes while a "user" could do this, in this context "user" really means "transcription developer".
There was a problem hiding this comment.
This would print out for every phase where this behavior is applicable. I agree that that's somewhat noisy, but users are capable of squelching it by explicitly setting the solver themselves.
| warn = True | ||
| phase.nonlinear_solver = om.NewtonSolver(iprint=0) | ||
| phase.nonlinear_solver.options['solve_subsystems'] = True | ||
| phase.nonlinear_solver.options['maxiter'] = 1000 |
There was a problem hiding this comment.
Why is maxiter being raised from 100 to 1000?
There was a problem hiding this comment.
I had seen one case where it was taking somewhat more than 100 iterations so I just bumped it up.
Summary
Previously dymos automatically assigned a DirectSolver to StateIndependentsComp when necessary, but this was problematic depending on the layout of the systems on MPI (mostly due to the presence of distributed systems within the ODE.
Dymos needs solvers on phases in the following conditions:
solve_segmentsis True for one or more states in pseudospectral and Birkhoff transcriptions.input_initialset toTruefor one or more statesduration_balanceis used and therefore a residual is associated with the time duration of a phase.By default, Dymos will automatically assign
NewtonSolverandDirectSolverto the phase if one of these conditions is met.Furthermore, if a Trajectory has its phases under a ParallelGroup and there are explicit connections between these phases, some form of solver is required to ensure that the inputs/ouputs between the phases are consistent.
By default, Dymos will add NonlinearBlockJac and PETScKrylov as the default solvers in
trajectory.phaseswhen this is the case.The user can bypass this behavior by assigning some other solver to the phase or
Trajectory.phases. If the user really truly wants to use a RunOnce solver in these situations, then the new Trajectory and Phase optionauto_solversshould be set toFalse.Adds new options to Trajectory:
parallel_phasesdictates whether the top level phase group is a ParallelGroup or a standard Group. Setting this toFalseallows the user to force phases to run in series when operating in parallel.auto_solversisTrueby default which causes dymos to attempt to place solvers where it thinks they are necessary. Setting this toFalsewill disable this behavior.New options for Phase
auto_solversisTrueby default which causes dymos to attempt to place solvers where it thinks they are necessary. Setting this toFalsewill disable this behavior.Related Issues
indep_states#1009Backwards incompatibilities
None
New Dependencies
None