Minerva can be reached via ssh to one of two login nodes, which are known by DNS as minerva01.aei.mpg.de and minerva02.aei.mpg.de. Internally, their names are login01 and login02.

• Learn how to type “minerva” without mixing the letters.
• Find the private counterpart of the submitted public key.
• Setup a couple of entries in your .ssh/config file, to make typing easier, and use the correct key.

Once logged in, you can access the other login node, and the compute nodes (node[001-594]) by passwordless ssh - a special keypair without passphrase has been created and installed for you. Do not use this key outside of Minerva.

• …
• …

Recommended SBATCH options (see also Slurm's Rosetta Stone)

...
...

Here's a script file stolen from CV:

#!/bin/bash

#SBATCH -J imb-03
#SBATCH -N 412
#SBATCH --ntasks-per-node 16
#SBATCH -t 20
#SBATCH -m cyclic:cyclic
#SBATCH -D /home/cvsupport/dmitry/benchmarks/imb/03
#SBATCH -o %N.%j.%a.out
#SBATCH -e %N.%j.%a.err

echo $SLURM_JOB_NODELIST

module purge
module load intel2016.3/mpi/64/5.1.3
export I_MPI_DEBUG=5
export I_MPI_FABRICS=shm:tmi
export I_MPI_FALLBACK=no
#export I_MPI_PIN=yes
#export I_MPI_PIN_PROCESSOR_LIST=8

mpirun -genvall -n $(($SLURM_JOB_NUM_NODES*$SLURM_NTASKS_PER_NODE)) -rr ../IMB-MPI1.intel2016.3 -multi 1 PingPing

• salloc -p devel … will submit a job to the selected partition
• if there are free resources (and you're allowed to use them!), this starts a new shell on the login node and assigns the requested resources
• you may check using env | grep SLURM

[stefgru@login01 ~]$ env | grep SLURM
SLURM_NODELIST=node579
SLURM_JOB_NAME=bash
SLURM_NODE_ALIASES=(null)
SLURM_NNODES=1
SLURM_JOBID=39809
SLURM_TASKS_PER_NODE=16
SLURM_JOB_ID=39809
SLURM_SUBMIT_DIR=/home/stefgru
SLURM_JOB_NODELIST=node579
SLURM_CLUSTER_NAME=minerva
SLURM_JOB_CPUS_PER_NODE=16
SLURM_SUBMIT_HOST=login01.cluster
SLURM_JOB_PARTITION=devel
SLURM_JOB_NUM_NODES=1

• you're still on te login node, srun will use the allocated node(s)
• (recommendation: start screen to have access to the node(s) in parallel)
• srun …
• exit will close the session, and return the allocated node(s) to the pool

It is recommended to use IntelMPI (IMPI) with Slurm (that's what PIK people say). Slurm and IMPI will do the correct binding for you, with OpenMPI you may see strange effects.

(to be extended)

Back to the main Minerva page