Non-reproducibility across layouts in tracer advection
Closed this issue · 3 comments
For the Bering domain, answers match for 21 vs. 24 cores, but not for 126 vs. 128 cores. The DEBUG output points to Post-advect temp and salt. This is with the PPM:H3 advection scheme. Changing the stencil to 3 in MOM_tracer_advect.F90 seems to fix it. Does this make sense? @Hallberg-NOAA
Trying the PPM scheme does not reproduce either. I see temperature differences at one spot near the open boundary, at the bottom.
The tracer reservoirs update with uhtr, vhtr, which are limited to two points of valid information in the halo. This is determined in lines mom-ocean#1033-1038 of MOM_dynamics_split.F90. If we need more points, we can add a pass_vector for tres_x, tres_y after they get updated in update_segment_tracer_reservoirs.