Recovering from checkpoint yields to different partitioning

marcfehling / hpbox

Sandbox for hp-adaptive methods

GNU General Public License v3.0

2 stars 1 forks source link

Recovering from checkpoint yields to different partitioning #14

Open marcfehling opened 1 month ago

marcfehling commented 1 month ago

Running the same cycle after successive refinements or after recovering from checkpoint results in different partitionings.

Try out:

check how the weighting signal is triggered at repartition()
call repartition() after execute_coarsening_and_refinement() to see if this gives the same partitioning

peterrum commented 1 month ago

Can you set: https://www.dealii.org/developer/doxygen/deal.II/classparallel_1_1distributed_1_1Triangulation.html#a0f5f69b2f422595604c9317ea7f15c13a92fad5ef4e0f7f69d008cb54123b1a9d?

marcfehling commented 1 month ago

I didn't know about this flag! I will try it out

marcfehling commented 1 month ago

I can't reproduce it with pure h-refinement. Only with hp-refinement and with weighted repartitioning and non-constant weights. I suspect the error is somewhere in the weighted load balancing part.

The mesh_reconstruction_after_repartitioning flag does not solve the issue.

marcfehling commented 1 month ago

Interesting. The flag mesh_reconstruction_after_repartitioning also drastically affects the total number of cells for consecutive refinements. With the flag, I have about 40% less cells after 5 hp-refinements in the Stokes scenario. That does not seem right.