How to improve the running speed of CESM2.0.1

[Mingjun]

Hi, everyone!

I am using the BHIST compset with a resolution of f19_g17. I have made many attempts to modify the pe-layout, but the fastest speed I could achieve is only 12~14 model years per day. The current "env_mach_pes.xml" file is attached. I aim to reach at least 50 years per day or more (because I need a very long period simulation). Is this feasible, and how should I modify the "env_mach_pes.xml" file to achieve my goal? Any suggestions would be helpful to me. Thank you!

Best regards.

 

[Mingjun]

Sorry, the right pe file is this one.

 

[jedwards]

5. Controlling processors and threads — CIME master documentation

You have all of the ROOTPES > 396 which means that pes 0-395 are doing nothing at all.
Look in file cesm/cime_config/config_pes.xml for some suggested pe layouts to start from.
The OCN and ATM should run concurrently and the ICE and LND should run sequentially with atm and
concurrently with OCN. So
NTASKS_ATM = NTASKS_ICE + NTASKS_LND
ROOTPE_OCN = NTASKS_ATM
ROOTPE_ICE=NTASKS_LND
NTASKS_ROF=NTASKS_LND
NTASKS_CPL=NTASKS_ATM
All other ROOTPE_*=0

once you get this basic configuration running you should fine tune it for your hardware.

 

[Mingjun]

5. Controlling processors and threads — CIME master documentation

You have all of the ROOTPES > 396 which means that pes 0-395 are doing nothing at all.
Look in file cesm/cime_config/config_pes.xml for some suggested pe layouts to start from.
The OCN and ATM should run concurrently and the ICE and LND should run sequentially with atm and
concurrently with OCN. So
NTASKS_ATM = NTASKS_ICE + NTASKS_LND
ROOTPE_OCN = NTASKS_ATM
ROOTPE_ICE=NTASKS_LND
NTASKS_ROF=NTASKS_LND
NTASKS_CPL=NTASKS_ATM
All other ROOTPE_*=0

once you get this basic configuration running you should fine tune it for your hardware.

Thank you very much for your response. I have referenced the official timing document and combined it with the principles you provided. Currently, my speed can reach 23 model years per day (although it hasn't reached my expectations yet, I will continue trying :)

 

[wfc1102@163_com]

Thank you very much for your response. I have referenced the official timing document and combined it with the principles you provided. Currently, my speed can reach 23 model years per day (although it hasn't reached my expectations yet, I will continue trying :)

Your MAX_TASKS_PER_NODE=56, I think NTASKS_??? and ROOTPE_??? = 56 * N, where N is integer, would save your computing resources and may speed up a little bit.