Repository navigation
Description
@s-poll noticed that OASIS spends around 60s on a 10-node eCLM-ICON runs, and this gets worse as the node count is increased (note the orange bars below).
oasis_send should be called as soon as all required fields are computed; similarly oasis_receive should be moved much closer to the subroutine that needs the fields. Currently, oasis_send/oasis_receive are invoked outside the main eCLM driver routine:
eCLM/src/clm5/cpl/lnd_comp_mct.F90
Lines 476 to 497 in 9c8f1ae
| #ifdef COUP_OAS_PFL | |
| call oas_receive(bounds, time_elapsed, atm2lnd_inst) | |
| #endif | |
| ! Run clm | |
| call t_barrierf('sync_clm_run1', mpicom) | |
| call t_startf ('clm_run') | |
| call t_startf ('shr_orb_decl') | |
| calday = get_curr_calday() | |
| call shr_orb_decl( calday , eccen, mvelpp, lambm0, obliqr, declin , eccf ) | |
| call shr_orb_decl( nextsw_cday, eccen, mvelpp, lambm0, obliqr, declinp1, eccf ) | |
| call t_stopf ('shr_orb_decl') | |
| call clm_drv(doalb, nextsw_cday, declinp1, declin, rstwr, nlend, rdate, rof_prognostic) | |
| call t_stopf ('clm_run') | |
| #if defined(COUP_OAS_ICON) | |
| call oas_send_icon(bounds, time_elapsed, lnd2atm_inst) | |
| #endif | |
| #if defined(COUP_OAS_PFL) | |
| call oas_send(bounds, time_elapsed, lnd2atm_inst) | |
| #endif |
Minimizing send/receive blocking is a good MPI practice; thus the OASIS calls need to be relocated to reduce coupling time. Non-blocking history file saving could further reduce OASIS delays, but this should be pursued separately and within the context of improving eCLM file I/O behavior.
Activity
kvrigor commentedon Oct 23, 2026
I'll start working on this after eCLM-ICON changes have been merged to
master. (#151).