Best practice -- Storage on Hawk

Change history:

Initial version ; Jose Gracia, 7 May 2024

TODOs:

Cleanup after parallel job: check if there is a HLRS recommendation to copy files from workspaces.

Available filesystems

$HOME

get current quota: $ na_quota
group quota: 200 GB, no files limit
user quota: 50 GB, no files limit
mounted via NFS; relatively slow

Workspaces

get current quota: $ ws_quota
group quota: 3 TB, 100k files
user quota: none
parallel file system Lustre; metadata slow, parallel access is fast
hitting quota limit, disables queues for whole group

/localscratch/$UID

total size: df -h /localscratch -> 22 TB
temporary scratch space
deleted at logout
local SSD; fast
available only on login nodes

What to put where?

Persistent data on $HOME. E.g:

source code
installed programs

Temporary data/builds on /localscratch/$UID. E.g:

anything temporary which does not fit on $HOME

Data for parallel jobs on workspace. E.g.

Input and outputs of jobs
anything which is access through MPI-IO or similar

Typical workflows

Large software project with autotools, e.g. OpenMPI

SRCFS=$HOME; BUILDFS=/localscratch/$UID; INSTALLFS=$HOME/opt

git clone --depth=1 git@github.com:open-mpi/ompi.git $SRCFS/ompi.git
cd $SRCFS/ompi.git; autoreconf -fiv

mkdir -p $INSTALLFS/ompi_test_347; mkdir -p $BUILDFS/build_ompi

cd $BUILDFS/build_ompi
$SRCFS/ompi.git/configure --prefix $INSTALLFS/ompi_test_347
make && make install

Large software project with CMake, e.g. targetDART

SRCFS=$HOME; BUILDFS=/localscratch/$UID; INSTALLFS=$HOME/opt

git clone --depth=1 git@github.com:targetDART/llvm-project.git $SRCFS/TD.git

mkdir -p $INSTALLFS/TD_buggy_again; mkdir -p $BUILDFS/build_TD

cd $BUILDFS/build_TD
cmake $SRCFS/TD.git/llvm Ninja -DCMAKE_INSTALL_PREFIX=$INSTALLFS/TD_buggy_again
make && make install

Avoiding large git checkouts

Do you really need all history? Do shallow clones.

git clone --depth=1 ...

Mirror git repo to $BUILDFS

Assuming you have a (possibly large) git repo my_repo.git on $HOME. Copy/clone source to $BUILDFS for faster access.

SRCFS=$HOME; BUILDFS=/localscratch/$UID; INSTALLFS=$HOME/opt

ls $SRCFS/my_repo.git
git clone --depth=1 file://$SRCFS/my_repo.git $BUILDFS/mirror_repo.git
mkdir $BUILDFS/build

cd $BUILDFS/build
cmake $BUILDFS/mirror_repo.git/ -DCMAKE_INSTALL_PREFIX=$INSTALLFS/whatever
$BUILDFS/mirror_repo.git/configure --prefix $INSTALLFS/whatever

Cleanup after parallel job

Rsynch results of parallel job into $HOME.

PERMANENTFS=$HOME/results/

cd $(ws_allocate my_job 1)
mpirun ./app --resultsdir=results
rsynch -a results $PERMANENTFS/
rm -rf results

2.9 KiB Raw Permalink Blame History