After the hybrid-assemble branch was merged into the trunk in May (at r3984) three of my parallel adaptive (multiphase flow) longtests started to silently fail (unfortunately there are currently no checks for solver convergence, and the tests still show as "passing" on buildbot even after the Velocity solve fails after just a few timesteps):
mphase_tephra_settling_2d_adaptive
mphase_tephra_settling_3d
mphase_tephra_settling_salinity_adaptive
I do not have OpenMP support enabled, and as far as I know the trunk does not configure with OpenMP either. These tests all run fine using the previous trunk revision (r3983), but each one fails after just a few timesteps with the hybrid-assemble contributions in r3984. The test I am currently looking at (mphase_tephra_settling_2d_adaptive) fails a few timesteps after the first mesh adapt. If mesh_adaptivity is disabled in the same simulation setup, everything runs fine and gives sensible results. All tests use the P1DG-P2 element pair and run over 8 processors.
After changing the following line in assemble/Momentum_DG.F90:
assemble_element = .not.dg.or.element_neighbour_owned(U, ele)
to:
assemble_element = .not.dg.or.element_owned(U, ele)
the mphase_tephra_settling_2d_adaptive test appears to run fine, but this isn't the correct thing to do. Any ideas?
On 10/08/12 16:15, Christian Jacobs wrote:
> Public bug reported:
...
> After changing the following line in assemble/
> assemble_element = .not.dg.
> to:
> assemble_element = .not.dg.
I was convinced at the time that element_
return .true. if ele was owned. i.e. it's a superset of
element_owned(U, ele). However, it's possible that for some corner
cases this may not be true.
Does it work if you change the test to:
.not.dg .or. (element_owned(U, ele) .or. element_
?
Lawrence