Setup a CESM 1.0.x benchmark run on a generic system

This is a cookbook to setup a CESM 1.0 benchmark run on a generic (Linux) system. This page is still work in progress, but will give already an idea what has to be done. See also porting CESM in the CESM user's guide: http://www.cesm.ucar.edu/models/cesm1.0/cesm/cesm_doc/c2161.html

System requirements

• Compilers known to work: intel 10.1, pgi 7.2, 8.0, 9.0, (pathscale 3.2)
• MPI implementations known to work: openmpi 1.4, 1.5, mvapich2 1.4, 1.5

Compile NETCDF (Requirement)

• See also http://www.ncl.ucar.edu/Download/build_from_src.shtml#NetCDF

• Download NETCDF

  wget http://www.unidata.ucar.edu/downloads/netcdf/ftp/netcdf-4.1.1.tar.gz
  tar xfvz netcdf-4.1.1.tar.gz
  cd netcdf-4.1.1
  

• Compile netcdf with the compiler you use later to compile the model. For example for intel compiler do:

  export FC=ifort
  export F77=ifort
  export F90=ifort
  export CPPFLAGS="-fPIC -DpgiFortran"
  ./configure --prefix=/usr/local/netcdf-4.1.1-intel --disable-netcdf-4 --disable-dap
  make
  make test
  

• Install NETCDF

  make install
  

Download CESM source code

• Download from NCAR SVN. A password is needed to checkout. Please register at: http://www.cesm.ucar.edu/models/cesm1.0/register/register_cesm1.0.cgi In case you don't have time to register ask urs.beyerle@env.ethz.ch

  svn export https://svn-ccsm-release.cgd.ucar.edu/model_versions/cesm1_0_2 cesm1_0_2
  

Adapt configuration files

• Change to Machines directory

  cd cesm1_0_2
  cd scripts/ccsm_utils/Machines/
  

• Meaning of filenames

  Filename	Purpose
  env_machopts.*	Set environment: Can be used to set paths to compiler, MPI library, NETCDF library
  Macros.*	Set compiler name and paths to MPI library, NETCDF library. Set compiler options
  mkbatch.*	Setting for queuing system

  where * corresponds to a machine.

• As starting point take configuration files of a machine that is close to your environment. For example have a look at brutus_io, brutus_im, brutus_po or brutus_pm where i=intel, p=pgi, o=openmpi, m=mvapich2

• Let's assume you use intel and openmpi, start with *.brutus_io files:

  cp env_machopts.brutus_io env_machopts.your_machine
  cp Macros.brutus_io       Macros.your_machine
  cp mkbatch.brutus_io      mkbatch.your_machine
  

• Add to config_machines.xml a configuration tag for your machine (your_machine) - only the important lines are listed below

  <machine MACH="your_machine"
           DESC="Test System"
           EXEROOT="/scratch/$CCSMUSER/$CASE"
           OBJROOT="$EXEROOT"
           INCROOT="$EXEROOT/lib/include" 
           DIN_LOC_ROOT_CSMDATA="/scratch/cesm1/inputdata"
           DIN_LOC_ROOT_CLMQIAN="/scratch/cesm1/inputdata/atm/datm7/atm_forcing.datm7.Qian.T62.c080727"
           BATCHQUERY="qstat -f"
           BATCHSUBMIT="qsub" 
           GMAKE_J="4" 
           MAX_TASKS_PER_NODE="4"
           MPISERIAL_SUPPORT="FALSE" />
  

  please set the following variables:

  EXEROOT=                     # working directory, final location of binary and output files
  DIN_LOC_ROOT_CSMDATA=        # input data, date will be downloaded on the fly
  DIN_LOC_ROOT_CLMQIAN=        # input data, data will be downloaded on the fly
  MAX_TASKS_PER_NODE=          # define cores per node
  

• Configure mpirun execution: Search in mkbatch.your_machine for the line starting the executable ccsm.exe and replace it with the correct mpirun command for your system, for example something like

  mpirun -np ${maxtasks} ./ccsm.exe >&! ccsm.log.\$LID
  # or
  mpirun -x LD_LIBRARY_PATH -np ${maxtasks} ./ccsm.exe >&! ccsm.log.\$LID
  

Compile and setup simulation

• Change to scripts directory

  cd cesm1_0_2
  cd scripts
  

• Define a case name (can be any name), for example 2° resolution (1.9x2.5_gx1v6), fully coupled model (B)

  CASE=1.9x2.5_gx1v6-B-benchmark
  

• Define the machine type, resolution, compset

  MACH=your_machine
  RES=1.9x2.5
  COMP=B
  

• Create case

  ./create_newcase -res $RES -compset $COMP -mach $MACH -case $CASE
  

• Change into case directory

  cd $CASE
  

• Define layout, for example run with 128 task on 128 cores

  NTASKS=128
  ./xmlchange -file env_mach_pes.xml -id NTASKS_ATM -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_LND -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_ICE -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_OCN -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_CPL -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_GLC -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id TOTALPES   -val $NTASKS
  

• In general, CESM is hardwired to generate monthly average data. In principle this can be turned of but needs a lot of code changes. Therefore it's not considered here. The following two cases are suggested instead:

• CASE 1: Run a short simulation with producing almost no output (I/O)
  Simulation should run only for 20 days and no restart files should be produced at the end:

  ./xmlchange -file env_run.xml -id STOP_OPTION -val ndays
  ./xmlchange -file env_run.xml -id STOP_N      -val 20
  ./xmlchange -file env_run.xml -id REST_OPTION -val never
  

• CASE 2: Run a larger simulation with producing monthly (better daily? - CKECK THIS) output data
  Simulation should run for 2 months and restart files should be produced at the end:

  ./xmlchange -file env_run.xml -id STOP_OPTION -val nmonths
  ./xmlchange -file env_run.xml -id STOP_N      -val 2
  ./xmlchange -file env_run.xml -id REST_OPTION -val $STOP_N
  

• Configure case

  ./configure -case
  

• Build/Compile the model

  ./$CASE.$MACH.build
  

Run the model

• Run the model, for example with LSF queuing system

  bsub < $CASE.$MACH.run
  

• To start without a queuing system just execute:

  ./$CASE.$MACH.run
  

• Timing results can be found after the run has been successfully completed in folder timing

  cat timing/ccsm_timing.$CASE.*
  ...
  Model Throughput:         6.39   simulated_years/day 
  ...
  

Change the layout

• To change the layout you don't have to recreate the case (but you can if you wish).
• Change into case directory and re-define layout

  cd $CASE
  NTASKS=64
  ./xmlchange -file env_mach_pes.xml -id NTASKS_ATM -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_LND -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_ICE -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_OCN -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_CPL -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id NTASKS_GLC -val $NTASKS
  ./xmlchange -file env_mach_pes.xml -id TOTALPES   -val $NTASKS
  

• Clean the case and re-configure it

  ./configure -cleanmach
  ./configure -case
  

• Build/Compile the model

  ./$CASE.$MACH.build
  

Change the resolution

• Recommended resolutions are T31_gx3v7 (~3°), 1.9x2.5_gx1v6 (2°), 0.9x1.25_gx1v6 (1°)
• To change the resolution create a new case !

Produce a summary

• Create performance matrix for CASE 1 and CASE 2. Fill in Model Throughput in simulated_years/day

• CASE 1: Run a short simulation with producing almost no output (I/O)

  resolution / layout (NTASKS)	16	32	64	128	256	512	1024 (a)
  T31_gx3v7					--	--	--
  1.9x2.5_gx1v6	--						--
  0.9x1.25_gx1v6	--	--

• CASE 2: Run a larger simulation with producing monthly (better daily? - CKECK THIS) output data

  resolution / layout (NTASKS)	16	32	64	128	256	512	1024 (a)
  T31_gx3v7					--	--	--
  1.9x2.5_gx1v6	--						--
  0.9x1.25_gx1v6	--	--

  (a) optional