{{{ #!html

Working on the Irene-amd machine

}}} ---- [[PageOutline(1-3,Table of contents,,numbered)]] # Connexion # {{{ ssh login@irene-amd-fr.ccc.cea.fr }}} # Introduction # * On-line users manual: https://www-tgcc.ccc.cea.fr/docs/irene.info.pdf (you will need a TGCC login and password) * Partition name : Rome * CPUs: 2x64 AMD Rome@2.6Ghz (AVX2) * !Core/Node: 128 * Nodes: 2292 * Total core: 293376 * RAM/Node: 256GB * RAM/core : 2GB * When submitting a job through ccc_msub or ccc_mprun, you must specify -m work, -m scratch, -m store, or combine them like in -m work,scratch. This constraint has the advantage that your jobs won't be suspended if a file system you don't need becomes unavailable. This is done in all jobs in libIGCM. * The default time limit for a job submission is 2 hours (7 200 s). * Irene post-processing nodes : xlarge are free and useful for post-processing operations. * Fat nodes for computation requiring a lot of shared memory (coming from irene.info) * Partition name: ''xlarge''' * CPUs: 4x28-cores Intel Skylake@2.1GHz * GPUs: 1x Nvidia Pascal P100 * !Cores/Node: 112 * Nodes: 5 * Total cores: 560 * RAM/Node: 3TB * RAM/Core: 5.3GB * IO: 2 HDD de 1 TB + 1 SSD 1600 GB/NVMe # Job manager commands # * {{{ccc_msub job}}} -> submit a job * {{{ccc_mdel ID}}} -> kill the job with the specified ID number * {{{ccc_mstat -u login}}} -> display all jobs submitted by login, add {{{-f}}} to see full job name * {{{ccc_mpp}}} -> display all jobs submitted on the machine. {{{ ccc_mpp -n }}} to avoid colors. * {{{ ccc_mpp -u $(whoami)}}} ->display your jobs. # Suggested environment # ## General environment ## Before working on Irene you need to prepare your environment. This is important to do before compilation to ensure the use of same modules as done by libIGCM running environment. We propose you 2 files which you can copy from the home igcmg. The first one called '''bashrc''' will source the second called '''bashrc_irene'''. Copy both files to your home, rename them by adding a dot as prefix. You can add personal settings in your .bashrc_irene-amd. Do as follow: {{{ cp ~igcmg/MachineEnvironment/irene-amd/bashrc ~/.bashrc # copy it even if you already have a .bashrc file cp ~igcmg/MachineEnvironment/irene-amd/bashrc_irene-amd ~/.bashrc_irene-amd }}} The .bashrc will source your own .bashrc_irene-amd which must be in your home. After re-connexion or source of .bashrc_irene-amd, check your loaded modules for intel, netcdf, mpi, hdf5 needed for the compilation: {{{ module list Currently Loaded Modulefiles: 1) ccc/1.0(default) 6) feature/mkl/lp64 11) c/intel/19.0.5.281(default) 16) feature/openmpi/mpi_compiler/intel(default) 21) sharp/2.0(default) 26) netcdf-c/4.6.0(default) 2) datadir/drf 7) feature/mkl/sequential 12) c++/intel/19.0.5.281(default) 17) flavor/openmpi/standard(default) 22) hcoll/4.4.2938(default) 27) netcdf-fortran/4.4.4(default) 3) datadir/own(default) 8) feature/mkl/single_node 13) fortran/intel/19.0.5.281(default) 18) feature/openmpi/net/auto(default) 23) ucx/1.7.0(default) 28) hdf5/1.8.20(default) 4) dfldatadir/own(default) 9) mkl/19.0.5.281 14) intel/19.0.5.281(default) 19) .tuning/openmpi/4.0.2 24) mpi/openmpi/4.0.2(default) 29) feature/bridge/heterogenous_mpmd 5) flavor/buildcompiler/intel/19 10) licsrv/intel 15) flavor/buildmpi/openmpi/4.0 20) hwloc/2.0.4 25) flavor/hdf5/parallel }}} The modules are specified in the file ~igcmg/MachineEnvironment/irene-amd/env_irene-amd which is sourced in bashrc_irene-amd. The same file env_irene-amd is sourced in libIGCM. --> Be careful this environment can be update during next weeks according to TGCC recommendations [[NoteBox(note, Create ~/.forward file in your main home containing only one line with your email address to receive emails from libIGCM. , 600px)]] ## Subversion version ## Since only recent subversion version (i.e > 1.6) are installed on Irene supercomputer, some usual functionalities are not available anymore (ex : svn command on copy of subdirectories...). In order to keep these functionalities, subversion 1.6.9 has been installed. To use this version : {{{ module unload subversion module use ~igcmg/Modules/tools module load subversion/1.6.9 irene190 : svn --version svn, version 1.6.9 (r901367) compiled Oct 31 2018, 11:12:49 }}} The use of this version has been added in default environment ~igcmg/MachineEnvironment/irene-amd/bashrc_irene-amd. Beware of coherence of subversion version you use from the extraction of your model/configuration to the use of svn commands in directories of your model/configuration. # How the storage project is set by libIGCM # When you use libIGCM it is recommended to dedicate one modipsl/libIGCM to one project allocation. By default, the output folders IGCM_OUT will be created in the directories $CCCSCRATCHDIR, $CCCWORKDIR and $CCCSTOREDIR corresponding to the project used in the main job. It is important that the same project is used in the post-processing jobs in libIGCM. [[NoteBox(warn,For gencmip6 project\, you have to set a subproject for computing and gencmip6 is forced for all directories., 600px)]] If you need to use another project for the computing than the storage, it is possible to set the variable '''!DataProject''' in config.card !UserChoices section, for example !DataProject=gen6328, read more [wiki:Doc/Setup#config.card here]. This project will be used for all output directories for the computing job and post-processing jobs even if they have another project for computing in the headers. The variable''' !DataProject''' can also be used if you work with different project allocations in the same modipsl. Only exception (harmless) is the first RUN_DIR folder which is always created in the $CCCSCRATCHDIR corresponding to the dfldatadir loaded in when submitting main job. When the job resubmits itself, the RUN_DIR will be in the same project space as the rest of the output. # Example of ins_job {{{ > ./ins_job ... Wait for the next question ... Hit Enter or give project ID (default is gen0000), possible projects are gen1111 gen2222 ... or other xxxcmip6 : aaacmip6 bbbcmip6 gen0000 (RETURN) ProjectID is gen0000 at Irene Hit Enter or give TYPE OF NODE required for post-processing (default is "xlarge"), possible types of nodes are "skylake" or "xlarge" : (RETURN) ProjectNode for post-processing is xlarge at Irene Hit Enter or give NUMBER OF CORES required for post-processing (default is "8") possible numbers of cores are "1" to "112" for xlarge : (RETURN) ProjectCore for post-processing is 8 Wait for the next question ... Hit Enter or give project ID (default is gen0000), possible projects are gen1111 gen2222 ... or other xxxcmip6 : aaacmip6 bbbcmip6 (RETURN) PostID is gen0000 at Irene on xlarge for post-processing }}} # File systems used on Irene # A figure to illustrate Irene filesystems is available [wiki:Doc/ComputingCenters/TGCC#TGCCsmachinesandfilesystems here] # Irene job headers # Here is an example of a job header as generated by libIGCM on the Irene-amd machine: {{{ ###################### ## IRENE TGCC/CEA ## ###################### #MSUB -r MY-SIMULATION #MSUB -o Script_Output_MY-SIMULATION.000001 #MSUB -e Script_Output_MY-SIMULATION.000001 #MSUB -eo #MSUB -n 976 #MSUB -x #MSUB -T 86400 #MSUB -A dekcmip6 #MSUB -q rome #MSUB -m store,work,scratch }}} The detail is as follows: || '''Control''' || '''Keyword''' || '''Argument''' || '''Example''' || '''Comments''' || || ''Job name'' || {{{-r}}} || string || {{{#MSUB -r Job_MY-SIMULATION}}} || The string should not contain the underscore (_) character. || || ''Standard output file name'' || {{{-o}}} || string || {{{#MSUB -o Script_Output_MY-SIMULATION.000001}}} || || || ''Error output file name'' || {{{-e}}} || string || {{{#MSUB -e Script_Output_MY-SIMULATION.000001}}} || If both {{{-o}}} and {{{-e}}} names are the same, the two outputs will be merged. The {{{-eo}}} option of the example is as a matter of fact redundant with that. || || ''Number of MPI tasks allocated'' || {{{-n}}} || integer || {{{#MSUB -n 976}}} || This is for the SPMD case; in the MPMD case, the integer is the number of cores. || || ''Node exclusivity'' || {{{-x}}} || none || {{{#MSUB -x}}} || || || ''Wall-time (maximum time allowed for execution)'' || {{{-T}}} || integer || {{{#MSUB -T 86400}}} || Time is expressed in number of seconds. || || ''Project allocation'' || {{{-A}}} || string || {{{#MSUB -A dekcmip6}}} || || || ''Partition used'' || {{{-q}}} || string || {{{#MSUB -q rome}}} || Choice is amongst {{{rome}}} || || ''Visible spaces'' || {{{-m}}} || (comma-separated) string(s) || {{{#MSUB -m store,work,scratch}}} || Specific to TGCC. The job can only "see" spaces specified on this line. Possible spaces are amongst {{{store}}} and/or {{{work}}} and/or {{{scratch}}}. || || ''Job priority'' || {{{-U}}} || string || {{{#MSUB -U high}}} || Possible arguments are {{{high}}}, {{{medium}}} or {{{low}}}. (This command does not appear by default on jobs generated by libIGCM.) || # Use of specific options to increase computing performances # Following options have to be used with caution. Please contact the person in charge of the configuration you use before launching production runs with these options. ## Increase the number of cores per MPI/OpenMP process ## Because of the architecture of Irene-amd processor, you could be limited by the memory available per core as well as the memory bandwidth. In order to use more cores per MPI process (or OpenMP thread), you can specify the factor you want to increase this number of cores in config.card by using '''DEP''' keyword as follows : {{{ ATM= (gcm.e, lmdz.x, 71MPI, 8OMP, 2DEP) SRF= ("" ,"" ) SBG= ("" ,"" ) OCE= (opa, opa.xx, 360MPI, 2DEP) ICE= ("" ,"" ) MBG= ("" ,"" ) CPL= ("", "" ) IOS= (xios_server.exe, xios.x, 12MPI, 2DEP) }}} In this example, the number of cores will be increased by a factor 2 for each component. ## Use of dedicated nodes for XIOS servers ## Some preliminary tests (bad performances) drive us to investigate the use of dedicated nodes for XIOS servers. To activate this functionality, you have to modify your config.card by using "DEDICATED" keyword as follows : {{{ ATM= (gcm.e, lmdz.x, 71MPI, 8OMP, 2DEP) SRF= ("" ,"" ) SBG= ("" ,"" ) OCE= (opa, opa.xx, 360MPI, 2DEP) ICE= ("" ,"" ) MBG= ("" ,"" ) CPL= ("", "" ) IOS= (xios_server.exe, xios.x, 12MPI,2DEP, DEDICATED) }}} Note that this functionality is only available for IOS component.