wiki:Doc/ComputingCenters/TGCC/IreneAmd

Version 1 (modified by aclsce, 22 months ago) (diff)

--

Working on the Irene-amd machine


1. 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

2. 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.

3. Suggested environment

3.1. 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/bashrc ~/.bashrc
cp ~igcmg/MachineEnvironment/irene/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

Create ~/.forward file in your main home containing only one line with your email address to receive emails from libIGCM.

3.2. 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.

4. 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.

For gencmip6 project, you have to set a subproject for computing and gencmip6 is forced for all directories.

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 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

5. File systems used on Irene

A figure to illustrate Irene filesystems is available here

6. 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.)