Purpose: To show how cpl5 interacts with the coupled system
(for eventual comparison with cpl6.)
Preliminary conclusion: For this processor configuration, cpl5 is not a bottleneck.
Please send any questions or comments to [email protected]
MPE_Log_event calls work like a timer. You make one call to start
the timer (start the color bar) and
one call to stop the timer (stop the color bar).
A typical application might look like this:
call MPE_Log_event( start_compute)
.....
......advect, do physics, etc.
....
call MPE_Log_event(end_compute)
call MPE_Log_event(start_comm)
call send_to_coupler
call MPE_Log_event(end_comm)
call MPE_Log_event(start_compute)
The various start_comm, start_compute variables are integers.
The result is a solid bar which changes color abruptly as the code
enters and leaves a state.
Three states of each model are shown: initialization, computation and communication.
Initialization for all 5 components is given the same color.
Computation is given a different color for each model: atmosphere, land, river, ocean, ice and coupler.
IMPORTANT: Communication is showing more than just the time for
MPI calls. It includes
wait time and whatever processing is done to form the message.
A key to what exactly is timed is below.
"Computation" includes everything but the communication routines listed below.
code used: CCSM2 May 17th release with Patch 1 applied
case: standard 5 day test run
machine: blackforest
processor config: default (68 nodes)
queue: csl_reg
Key: 1 bar for each MPI process
Top 8 bars: atmosphere
Next 3 bars: land and river
Next 16 bars: ice
Next 40 bars: ocean
bottom bar: coupler
Communication routines timed (same color used for each side of communication).
c2o: msg_ocns (cpl5), recv_from_coupler (pop)
o2c: msg_ocnr (cpl5), send_to_coupler
(pop)
c2l: msg_lnds (cpl5), csm_recv (clm2)
l2c: msg_lndr (cpl5), csm_send (clm2)
c2i: msg_ices (cpl5), from_coupler (csim4)
i2c: msg_icer (cpl5), to_coupler (csim4)
c2a: msg_atms (cpl5), ccsmrcv (cam)
a2c: msg_atmr (cpl5), ccsmsnd (cam)
More info: