Updating step size control #483

hadjimy · 2014-12-16T12:43:17Z

This PR changes the step-size control for LMMs and revises the way the step-size is updated at each step for RK methods.

There is still some work currently in process to make the code more efficient and robust.

Note there is a test that fails in parallel (acoustics_2d_homogeneous) when SSPLMM32 is used, but could not reproduce that in my machine (where all tests pass).

…ze calculatiob.

…ent solution.

…ssic. Remove step_index reset to sharpclaw solver.

…. step_index is updated only when step is accepted.

Also: - Change LMM names from 'MS' to 'LMM' - Make step rejection work correctly for LMMs - Allow use of grid norms in util.check_diff() by passing delta

Due to step rejection occurring since we now check the CFL number for all RK stages.

…tore them. Also improve the way stored information from previous steps is updated.

…lobal cfl at the beggining of each step.

mandli · 2014-12-17T22:55:53Z

Any ideas on the test failure front? It would be nice to merge this in but I really don't think we should if the tests fail even if it is not reproducible.

hadjimy · 2014-12-18T08:14:13Z

No it is unclear why this single test fails. I run nosetests on both my mac and Ubuntu with no errors.
@mandli Could you run nosetests in your machine?
We shouldn't merge it before figuring out was it going wrong.
Note that SSPLMM32 is used in acoustics_2d_variable (and other examples) with no errors.

mandli · 2014-12-20T23:03:15Z

The tests pass although PETSc does emit an error I have not seen before

[0]PETSC ERROR: Caught signal number 11 SEGV: Segmentation Violation, probably memory access out of range
[0]PETSC ERROR: Try option -start_in_debugger or -on_error_attach_debugger
[0]PETSC ERROR: or see http://www.mcs.anl.gov/petsc/documentation/faq.html#valgrind[0]PETSC ERROR: or try http://valgrind.org on GNU/linux and Apple Mac OS X to find memory corruption errors
[0]PETSC ERROR: configure using --with-debugging=yes, recompile, link, and run 
[0]PETSC ERROR: to get more information on the crash.

I am not sure if that's anything to do with the problem you are seeing as I cannot figure out which test is actually emitting the error and why it does not cause the test to fail.

ketch · 2014-12-21T13:03:58Z

@mandli I believe you are running nosetests from clawpack/pyclaw. Run it from clawpack/pyclaw/examples. See #489.

mandli · 2014-12-21T20:14:01Z

@ketch thanks for tracking that one down.

hadjimy · 2014-12-22T06:26:08Z

I can reproduce Travis error message when running
mpirun -n 4 nosetests:

======================================================================
ERROR: examples.acoustics_2d_homogeneous.acoustics_2d({'time_integrator': 'SSPLMM32', 'disable_output': True, 'solver_type': 'sharpclaw'})
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/hadjimy/Library/Enthought/Canopy_64bit/User/lib/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "/Users/hadjimy/Programs/clawpack/clawpack/pyclaw/util.py", line 172, in <lambda>
    test = lambda: test_app(application, verifier, test_kwargs)
  File "/Users/hadjimy/Programs/clawpack/clawpack/pyclaw/util.py", line 258, in test_app
    raise VerifyError(err)
VerifyError: /Users/hadjimy/Programs/clawpack/pyclaw/examples/acoustics_2d_homogeneous/acoustics_2d.py
********************************************************************************
verification function
        def verify(claw):
            """ verifies 2d homogeneous acoustics from a previously verified run """
            import os
            import numpy as np
            from clawpack.pyclaw.util import check_diff


            #grabs parallel results to process 0, None to other processes
            test_q=claw.solution.state.get_q_global()

            if test_q is not None:
                test_pressure = test_q[0,:,:]
                thisdir = os.path.dirname(__file__)
                expected_pressure = np.loadtxt(os.path.join(thisdir,data_filename))
                return check_diff(expected_pressure, test_pressure, reltol=1e-3, 
                                    delta=claw.solution.grid.delta)
            else:
                return

args                 : {'use_petsc': True, 'time_integrator': 'SSPLMM32', 'kernel_language': 'Fortran', 'disable_output': True, 'solver_type': 'sharpclaw'}
norm of expected data: 0.0196003663743
norm of test data    : 0.0195695757644
test error           : 0.000185196027373
reltol  : 0.001
********************************************************************************

-------------------- >> begin captured stdout << ---------------------
{'use_petsc': True, 'time_integrator': 'SSPLMM32', 'kernel_language': 'Fortran', 'disable_output': True, 'solver_type': 'sharpclaw'}

--------------------- >> end captured stdout << ----------------------
-------------------- >> begin captured logging << --------------------
root: INFO: Solution 0 computed for time t=0.000000
root: INFO: Solution 1 computed for time t=0.012000
root: INFO: Solution 2 computed for time t=0.024000
root: INFO: Solution 3 computed for time t=0.036000
root: INFO: Solution 4 computed for time t=0.048000
root: INFO: Solution 5 computed for time t=0.060000
root: INFO: Solution 6 computed for time t=0.072000
root: INFO: Solution 7 computed for time t=0.084000
root: INFO: Solution 8 computed for time t=0.096000
root: INFO: Solution 9 computed for time t=0.108000
root: INFO: Solution 10 computed for time t=0.120000
--------------------- >> end captured logging << ---------------------

Before I was just running nosetests and as @ketch pointed out this does not run any tests in parallel.
I made another branch and tried a bit to see what is going wrong.

In my opinion this error is not caused by the recent changes in step-size control. It is an error related to TVD discretizations (lim_type=1) and I think it has been around for some time.
Note that there are only 3 nosetests to test the lim_type=1 option:

in euler 1d which runs ok
in acoustics 2d variable
in acoustics 2d homogeneous

The error occurs when testing sharpclaw with TVD methods (solver_type=sharpclaw and lim_type=1) because the difference between the computed and reference solution is large (eg. when measured in L1 grid norm). This is not noticeable in acoustics_2d_variable because nosetests check the absolute error with a tolerance 1e-1. However in acoustics 2d_homogeneous nosetests check the relative error and this triggers the error when running in MPI and PETSc.

You can test it out by checking nosetests_petsc from
https://github.com/hadjimy/pyclaw/tree/nosetests_petsc

Then run checking_petsc.sh and observe the error between the expected_pressure and test_pressure when running with WENO and TVD methods. That should have been of order machine precision but when lim_type=1 and use MPI and PETSc, the error is more than 1e-4.

This might be related to errors accumulated at ghost cells or might be more involved and has to do with how TVD methods work in MPI and PETSc.

ketch · 2014-12-30T09:21:36Z

src/pyclaw/cfl.py

@@ -22,6 +22,8 @@ def get_cached_max(self):
    def set_local_max(self,new_local_max):
        self._global_max = new_local_max

-    def update_global_max(self,new_local_max):
-        self._global_max = new_local_max
+    def set_global_max(self,new_global_max):


I think this method is identical to that of the parent function, so there is no need to explicitly duplicate it here. It will be automatically inherited.

Correct. I removed that method. Also set_local_max is not currently being used.

hadjimy · 2015-01-11T06:50:19Z

@ketch Do you think we should test LMM only with WENO and merge this PR?

Add set_global back to CFL object. Only update dt_list and dtFE_list when a step is accepted.

…eject_step function.

…eject_step function (2).

mandli · 2015-02-17T20:39:14Z

Excellent, I wanted to make sure as we had talked about making this change before. I would say we should merge this then unless someone else has objections.

hadjimy · 2015-02-17T20:43:24Z

There is some space for improvements, for example using the already saved function evaluations for a general LMM method and I recently discovered that we are using an additional register for the SSP104 method. But these are smaller changes and can be included in a forthcoming PR.

mandli · 2015-02-17T20:44:35Z

Yeah, I would say merge this and add those as upcoming enhancements. There are a few spots that I think could be PEP8ized as well but that's pretty minor.

ketch · 2015-02-19T06:51:18Z

Thanks, @mandli. @hadjimy is investigating some funny observed behavior of the initial step sizes before we merge this.

Also: - General LMM now uses previously saved function evaluations - rename self.deltaq to self.dq_dt - update accept_reject_step and get_dt_new

hadjimy · 2015-02-23T09:09:46Z

I have fixed the issue regarding the step-size choice for starting methods and also make general LMM use previous function evaluations.

…nto step-size_control * 'step-size_control' of git://github.com/hadjimy/pyclaw: Update nosetests. Correct step-size choice for starting methods of LMMs. Conflicts: src/pyclaw/sharpclaw/solver.py

This avoids some messy things and ensures that the step size rejection is determined using the right CFL number.

Also: - Add SSPLMM53 - Edit documentation

- Correct omega3 in SSPLMM43. - Add SSP coefficients for 3rd order SSPLMM. - Don't pass cfl to accept_reject_step(). - Make accept_step local variable in accept_reject_step(). - Use [:] instead of copy. - Edit documentation/remove redundant comments.

with the desired order.

self.step(). - Make SSPLMM work with constant step size.

- Make a single routine to update saved info for LMMs at the beggining of step() function. - Step size for starting values is chosen inside get_dt_new() and when the linear multistep part is active, the step size is set in set_dt(). - Add routine for the additional conditions for 3rd order SSPLMM and correct step-size update in case the conditions are violated.

coveralls · 2015-03-25T13:27:42Z

Coverage increased (+0.44%) to 27.2% when pulling eb7d715 on hadjimy:step-size_control into 40a7249 on clawpack:master.

coveralls · 2015-03-26T08:38:35Z

Coverage increased (+0.71%) to 27.47% when pulling deb71d0 on hadjimy:step-size_control into 40a7249 on clawpack:master.

coveralls · 2015-03-26T13:06:31Z

Coverage increased (+0.72%) to 27.48% when pulling 7dfd2ed on hadjimy:step-size_control into 40a7249 on clawpack:master.

ketch · 2015-03-26T13:35:40Z

This looks ready to merge. It is great that now the 2nd-order methods allow for any number of steps and that there are tests for 4 different SSP LMMs. Nice work @hadjimy !

Updating step size control

ketch · 2015-03-26T13:36:45Z

@mandli and @rjleveque PyClaw is ready for the 5.3 release. There is a small additional refactoring that we hope to get done in time, but if you want to generate a release candidate now, that is fine.

hadjimy and others added 14 commits December 16, 2014 12:57

Update variable step-size SSPLMM.

ab35bbc

Make a function to accept/reject step and update code for new step-si…

ffed0f7

…ze calculatiob.

Update pyclaw solver to match with changes in sharpclaw solver.

f59a1b9

Edit how cfl number is updated, so cfl at each step relates with curr…

ab16359

…ent solution.

Reset local cfl at the beggining of step in classic.

493c141

Add functions for accept/reject step and new step-size choice for cla…

6d5b5fd

…ssic. Remove step_index reset to sharpclaw solver.

Major edits/corrections in accept/reject step and new step-size setup…

732d99a

…. step_index is updated only when step is accepted.

Correct step-size update at the last RK step.

046e929

Improve the way information from previous steps is updated.

9f6571d

Make LMM tests pass.

93ab5dd

Also: - Change LMM names from 'MS' to 'LMM' - Make step rejection work correctly for LMMs - Allow use of grid norms in util.check_diff() by passing delta

Update regression data for WC blast test.

e97bcfe

Due to step rejection occurring since we now check the CFL number for all RK stages.

Tests passing.

52711c0

Remove stored dt and dtFE from registers and create solver lists to s…

24a1c94

…tore them. Also improve the way stored information from previous steps is updated.

Add a function to set cfl._global_max and correct initialization of g…

419a068

…lobal cfl at the beggining of each step.

hadjimy force-pushed the step-size_control branch from cb9869f to 419a068 Compare December 17, 2014 10:58

ketch mentioned this pull request Dec 22, 2014

SharpClaw with TVD limiting gives different results in serial and parallel #491

Open

ketch reviewed Dec 30, 2014
View reviewed changes

Remove 'set_global_max' method from cfl.py

5607c44

ketch and others added 5 commits January 11, 2015 12:50

Refactoring the time integration code.

9a31120

Add set_global back to CFL object. Only update dt_list and dtFE_list when a step is accepted.

Debugging LMM startup.

798531e

Cleaning pyclaw solver and moving updating of stored info to accept_r…

c4e1bda

…eject_step function.

Cleaning pyclaw solver and moving updating of stored info to accept_r…

573f726

…eject_step function (2).

Use status['numsteps'] instead of step_index.

315e4b9

hadjimy added 2 commits February 23, 2015 10:09

Correct step-size choice for starting methods of LMMs.

2a72e9a

Also: - General LMM now uses previously saved function evaluations - rename self.deltaq to self.dq_dt - update accept_reject_step and get_dt_new

Update nosetests.

96223d5

ketch and others added 13 commits February 24, 2015 14:34

Evaluate dq at beginning of step.

bf3b73e

This avoids some messy things and ensures that the step size rejection is determined using the right CFL number.

Prevent LMM integrators from going past tend.

683061e

Correct dtFE computation for SSPLMM

2151691

Also: - Add SSPLMM53 - Edit documentation

Move section that saves self.dq_dt list so that SSPLMM43 converges

ed93f74

with the desired order.

- Move additional conditions for 3rd order SSPLMM at the beginning of

da96dc4

self.step(). - Make SSPLMM work with constant step size.

Make nosetests run.

027ed22

combine SSPLMM43 and SSPLMM53 code

2dd9d41

Combine code for all SSPLMM methods.

5937ea2

clean up comments and reorder if statements in sharpclaw/solver.py

d6dd638

move step rewinding up into update_saved_values

eb7d715

hadjimy added 2 commits March 26, 2015 11:23

Make SSPLMMs generic for arbitrary number of steps

3095391

Update nosetests: include tests for 3rd-order SSPLMM.

deb71d0

use solver.lmm_steps as definitive value

7dfd2ed

ketch added a commit that referenced this pull request Mar 26, 2015

Merge pull request #483 from hadjimy/step-size_control

e34ec5f

Updating step size control

ketch merged commit e34ec5f into clawpack:master Mar 26, 2015

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Updating step size control #483

Updating step size control #483

hadjimy commented Dec 16, 2014

mandli commented Dec 17, 2014

hadjimy commented Dec 18, 2014

mandli commented Dec 20, 2014

ketch commented Dec 21, 2014

mandli commented Dec 21, 2014

hadjimy commented Dec 22, 2014

ketch Dec 30, 2014

hadjimy Jan 11, 2015

hadjimy commented Jan 11, 2015

mandli commented Feb 17, 2015

hadjimy commented Feb 17, 2015

mandli commented Feb 17, 2015

ketch commented Feb 19, 2015

hadjimy commented Feb 23, 2015

coveralls commented Mar 25, 2015

coveralls commented Mar 26, 2015

coveralls commented Mar 26, 2015

ketch commented Mar 26, 2015

ketch commented Mar 26, 2015

Updating step size control #483

Updating step size control #483

Conversation

hadjimy commented Dec 16, 2014

mandli commented Dec 17, 2014

hadjimy commented Dec 18, 2014

mandli commented Dec 20, 2014

ketch commented Dec 21, 2014

mandli commented Dec 21, 2014

hadjimy commented Dec 22, 2014

ketch Dec 30, 2014

Choose a reason for hiding this comment

hadjimy Jan 11, 2015

Choose a reason for hiding this comment

hadjimy commented Jan 11, 2015

mandli commented Feb 17, 2015

hadjimy commented Feb 17, 2015

mandli commented Feb 17, 2015

ketch commented Feb 19, 2015

hadjimy commented Feb 23, 2015

coveralls commented Mar 25, 2015

coveralls commented Mar 26, 2015

coveralls commented Mar 26, 2015

ketch commented Mar 26, 2015

ketch commented Mar 26, 2015