Skip to content

Latest commit

 

History

History
87 lines (56 loc) · 2.92 KB

README.md

File metadata and controls

87 lines (56 loc) · 2.92 KB
Raw

Preparation of Abel before using the images framework:

  1. add module purge to /etc/bash.bash_logout on all compute nodes This allows module cleanup when user uses an interactive login to a compute node, or a login node.

  2. add to /etc/sudoers.d/users on all compute nodes

%users ALL=(ALL) NOPASSWD: /cluster/bin/mount_image %users ALL=(ALL) NOPASSWD: /cluster/bin/umount_image

Change /etc/sudoers to allow login without a tty

sed -i -e 's/^Defaults.*requiretty/# Defaults requiretty/' /etc/sudoers

  1. make sure that the path to tools inside module_load is correct:

TOOLS_PATH="/cluster/bin"

  1. If doing automatic migration, in update_module_scripts.sh, set path to module_load

/cluster/bin/module_load

and as root run update_module_scripts.sh

If converting a single module, the following line needs to be added at the end of the module file (next to last line) to enable image mounting:

system /cluster/bin/module_load $appname/$appversion ${SLURM_JOB_ID=NOJOBID} $action | logger -t software_images

  1. add slurm TaskProlog and JobEpilog that clears modules for a given job, or does a module purge before the job is done

/hpc/src/slurm/scripts/prolog_task: append

Mount software images:

if [[ -e /cluster/etc/use_software_images ]]; then ## Debugging: mfa=type -t module if [[ $mfa == "" ]]; then logger -t software_images_prolog $SLURM_JOB_ID hostname "module command not availabe (II), printing environment" set > /tmp/$SLURM_JOB_ID.II.set fi

modules=`module list --terse 2>&1 | grep -v "Currently Loaded"`
for m in $modules; do
    if [[ ! -e /cluster/etc/modulefiles/$m ]]; then
        logger -t software_images_prolog $SLURM_JOB_ID $m: unknown module
        continue
    fi
    # logger -t software_images_prolog $SLURM_JOB_ID `hostname` checking $m
    using_modules=`cat /cluster/etc/modulefiles/$m | grep /cluster/bin/module_load`
    if [[ $using_modules == "" ]]; then
        logger -t software_images_prolog $SLURM_JOB_ID `hostname` module $m NOT imaged
        continue
    fi
    logger -t software_images_prolog $SLURM_JOB_ID `hostname` loading image $m
    /cluster/bin/module_load $m $SLURM_JOB_ID load 2>&1 | logger -t software_images
done

fi

/hpc/src/slurm/scripts/epilog_slurmd: append

Umount software images:

if [[ -e /cluster/etc/use_software_images ]]; then logger -t software_images_epilog hostname unmounting all images for job $SLURM_JOB_ID /cluster/bin/umount_all_images --job_id $SLURM_JOB_ID 2>&1 | logger -t software_images if (( $? != 0 )); then remove_log=0 fi fi

-------------- ISSUES

double-cache issue. loopback device caches blocks, FS caches files. fixed in linux 4.4... http://unix.stackexchange.com/questions/278647/overhead-of-using-loop-mounted-images-under-linux

similar when dealing with containers https://lwn.net/Articles/588309/

ploop can solve this?