Question about static refinement.

[ToguyC]

Hello !

Thank you in advance for your time.

In order to better understand mesh refinement, I tried applying static mesh refinement on the simple heat example. I have chosen to use 2 levels of refinement with a refinement ratio at 2. I have set the finer grid to span the right side only (x > 0.5) and moved the Gaussian for the heat initialization on the left side (x < 0.5) in order to see the transfer from coarse to fine grid.

I understand that the coarse level should span on the whole domain in order to have a proper coordinate setup, but here is a point that I do not really understand.

When plotting using yt and matplotlib, I see that the coarse and fine levels holds values even for the cells that should not be handled by the respective level, i.e. cells in the fine grid holds values even though they are located on the left side (coarse).

I think that for the fine grid it must be because of the average_down, which copies values from the coarser level to all the other cells.

Also, currently I have my coarse grid spanning over the entire domain, i.e., not stopping at the half point, but when I tried to make it stop at the half point, everything broke.

Could you please help me better understand the behavior of the FillPatchSingleLevel and FillPatchTwoLevels, in addition with the average_down ?

Sorry to bother you with this,
Thank you in advance for your help,

Tanguy

[ToguyC]

It might be entirely caused by how yt retrieve data. Just to be sure.

[WeiqunZhang]

In AMReX, the level 0 grids usually cover the entire domain. So having coarse cells on the right half is expected. (Note that AMReX's grid is not a tree. A tree is just a special case for us. So generally speaking it's not possible to not allocate coarse cells covered by fine cells.) Having fine cells on the left half might be a yt thing for visualization purpose. As for the behavior of FillPatchSingleLevel, FillPatchTwoLevels and average_down, do you have specific questions? Do these Doxygen documentations answer your question?

https://amrex-codes.github.io/amrex/doxygen/namespaceamrex.html#a3cf973090fa613c635340a08df7ecd38

https://amrex-codes.github.io/amrex/doxygen/namespaceamrex.html#a0c7ffe6f08d1495383e8b7a360f5067e

https://amrex-codes.github.io/amrex/doxygen/namespaceamrex.html#a5eed60d257307d60b712f3d5ffd7cf49

[ToguyC]

Ok thank you for the precision. I figured it was a yt thing.

I kinda understand the behavior now yes, but I still have a question in mined regarding the update "rules" for multi level setup.
Earlier in the day, I just applied the phi update to all cells, regardless if a cell was covered by a finer grid or not.

But now, I managed to mask and skip the cells covered by a finer grid, and the results do not change (expected).

Which is better by the AMReX framework ? Is a way preferred from another ?

[WeiqunZhang]

First, it depends on your algorithm. Oftentimes, you have to include every coarse cells including the ones covered by fine cells. For example, you might have subcyling in time and you have elliptic equations to solve.

Second, even if you can get away with not advancing covered coarse cells, you may not gain much in performance, because the number of cells involved is only 1/8 of the number of fine cells in 3D. It could even be slower because of the extra mask you have to use.