Reproductibility issue with nn_comm=1 + flux form momentum advection schemes using small mpp domain sizes
ORCA2_ICE_PISCES configuration (with default namelists) is not reproducible when using nn_comm = 1
and ln_dynadv_cen2
or ln_dynadv_ubs
advection schemes when using ~200 processors. Switching to nn_comm = 2
solves the issue.
Tested so far:
- mpp splitting:
jpni = 8, jpnj = 4, jpnij = 32
vsjpni=14, jpnj=19, jpnij=223
- ORCA2_ICE_PISCES is reproducible when using
ln_dynadv_vec
or less processors - Land suppression is not to blame
- Issue is also present with
key_linssh
- no change with
ln_nnogather = .false.
- no more repro problem with
nn_comm = 2
- repro issue is present on multiple clusters (belenos & kara)
Edited by Guillaume Samson