Both the iterative wind solver and the wind rotation to the grid appear to have minor errors that change the windfield when run in parallel relative to the value when run in serial. Setting a larger halo size (3) fixes this, but with a very large computational cost.
Expected Behavior
wind fields are the same in serial and parallel
Current Behavior
difference between serial and parallel wind fields appear between processes.
Possible Solution
increase size of u and v grids to provide more information outside of immediate boundary, without increasing the halo being passed in all variables.
Steps to Reproduce (for bugs)
cafrun -np 36 icar
cafrun -np 1 icar
difference output files
Your Environment
Version of the code used: latest develop (pre 2.1 release)
Operating System and version: linux
Compiler and version: gfortran 10.1 opencoarrays 2.9.2
scrasmussen
commented
1 year ago
Both the iterative wind solver and the wind rotation to the grid appear to have minor errors that change the windfield when run in parallel relative to the value when run in serial. Setting a larger halo size (3) fixes this, but with a very large computational cost.
Expected Behavior
wind fields are the same in serial and parallel
Current Behavior
difference between serial and parallel wind fields appear between processes.
Possible Solution
increase size of u and v grids to provide more information outside of immediate boundary, without increasing the halo being passed in all variables.
Steps to Reproduce (for bugs)
Your Environment