Hi everyone,
I’m running CESM2 B1850cmip6 piControl. I ran ~12 model years with one PE layout, then continued from the same restart but changed the PE layout (9 nodes → 18 nodes; total cores ~574, different decomposition).
As a quick test, starting from the same restart I ran ~3 months with both layouts and the patterns diverge noticeably.
Is this expected (round-off + chaos)? More importantly, is it scientifically acceptable to change PE layout mid-run for a long piControl, or could it bias long-term statistics such as ENSO (Niño3.4 variance/spectrum) beyond losing bit-for-bit reproducibility?
I’m running CESM2 B1850cmip6 piControl. I ran ~12 model years with one PE layout, then continued from the same restart but changed the PE layout (9 nodes → 18 nodes; total cores ~574, different decomposition).
As a quick test, starting from the same restart I ran ~3 months with both layouts and the patterns diverge noticeably.
Is this expected (round-off + chaos)? More importantly, is it scientifically acceptable to change PE layout mid-run for a long piControl, or could it bias long-term statistics such as ENSO (Niño3.4 variance/spectrum) beyond losing bit-for-bit reproducibility?