Main menu

Navigation

How to create an optimum PE layout for B1850C5 compset.

1 post / 0 new
midhun.ndr@...
How to create an optimum PE layout for B1850C5 compset.
Hi,  I am running CESM 1.2 with B1850C5 compset. I have 12 nodes available with 20 cores per node (http://www.nlhpc.cl/es/cluster2014/) making a total of 240 cores.Currently, I make simulations by giving 240 cores to each components.   Here is my PE layout.<entry id="NTASKS_ATM" value="240"/>
<entry id="NTHRDS_ATM" value="1"/>
<entry id="ROOTPE_ATM" value="0"/>
<entry id="NINST_ATM" value="1"/>
<entry id="NINST_ATM_LAYOUT" value="concurrent"/>
<entry id="NTASKS_LND" value="240"/>
<entry id="NTHRDS_LND" value="1"/>
<entry id="ROOTPE_LND" value="0"/>
<entry id="NINST_LND" value="1"/>
<entry id="NINST_LND_LAYOUT" value="concurrent"/>
<entry id="NTASKS_ICE" value="240"/>
<entry id="NTHRDS_ICE" value="1"/>
<entry id="ROOTPE_ICE" value="0"/>
<entry id="NINST_ICE" value="1"/>
<entry id="NINST_ICE_LAYOUT" value="concurrent"/>
<entry id="NTASKS_OCN" value="240"/>
<entry id="NTHRDS_OCN" value="1"/>
<entry id="ROOTPE_OCN" value="0"/>
<entry id="NINST_OCN" value="1"/>
<entry id="NINST_OCN_LAYOUT" value="concurrent"/>
<entry id="NTASKS_CPL" value="240"/>
<entry id="NTHRDS_CPL" value="1"/>
<entry id="ROOTPE_CPL" value="0"/>
<entry id="NTASKS_GLC" value="240"/>
<entry id="NTHRDS_GLC" value="1"/>
<entry id="ROOTPE_GLC" value="0"/>
<entry id="NINST_GLC" value="1"/>
<entry id="NINST_GLC_LAYOUT" value="concurrent"/>}
<entry id="NTASKS_ROF" value="240"/>
<entry id="NTHRDS_ROF" value="1"/>
<entry id="ROOTPE_ROF" value="0"/>
<entry id="NINST_ROF" value="1"/>
<entry id="NINST_ROF_LAYOUT" value="concurrent"/>
<entry id="NTASKS_WAV" value="240"/>
<entry id="NTHRDS_WAV" value="1"/>
<entry id="ROOTPE_WAV" value="0"/>
<entry id="NINST_WAV" value="1"/>
<entry id="NINST_WAV_LAYOUT" value="concurrent"/>
<entry id="PSTRID_ATM" value="1"/>
<entry id="PSTRID_LND" value="1"/>
<entry id="PSTRID_ICE" value="1"/>
<entry id="PSTRID_OCN" value="1"/>
<entry id="PSTRID_CPL" value="1"/>
<entry id="PSTRID_GLC" value="1"/>
<entry id="PSTRID_ROF" value="1"/>
<entry id="PSTRID_WAV" value="1"/>
<entry id="TOTALPES" value="240"/>
<entry id="PES_LEVEL" value="2rp"/>
<entry id="MAX_TASKS_PER_NODE" value="20"/>
<entry id="PES_PER_NODE" value="20"/>
<entry id="COST_PES" value="0"/>
<entry id="CCSM_PCOST" value="-1"/>
<entry id="CCSM_TCOST" value="0"/>
<entry id="CCSM_ESTCOST" value="3"/></config_definition>


Overall I get a modelthroughput of 6.4myears/wday.  Here is the component wise breakup.  
    TOT Run Time:   13483.118 seconds       36.940 seconds/mday         6.41 myears/wday
    LND Run Time:      78.358 seconds        0.215 seconds/mday      1102.63 myears/wday
    ROF Run Time:       7.841 seconds        0.021 seconds/mday     11019.00 myears/wday
    ICE Run Time:    1838.970 seconds        5.038 seconds/mday        46.98 myears/wday
    ATM Run Time:    6411.420 seconds       17.566 seconds/mday        13.48 myears/wday
    OCN Run Time:    3186.190 seconds        8.729 seconds/mday        27.12 myears/wday
    GLC Run Time:       0.000 seconds        0.000 seconds/mday         0.00 myears/wday
    WAV Run Time:       0.000 seconds        0.000 seconds/mday         0.00 myears/wday
    CPL Run Time:     963.904 seconds        2.641 seconds/mday        89.64 myears/wday

How can I optimize the model throughput?
CAM5 is the most expensive component in the compset, but total model throughput is about only half of CAM5 throughput. Why is it so?

Who's new

  • kimmo.ruosteeno...
  • wangna@...
  • julio.fernandez@...
  • fjaume@...
  • sha59@...