Piz Daint

Piz Daint uses SLURM for the submission, monitoring and control of parallel jobs. Parallel programs compiled with Cray-MPICH (the MPI library available on this system) must be run using the srun command. SLURM batch scripts need to be submitted with the sbatch command from the $SCRATCH folder, since users are NOT supposed to run jobs from different filesystems because of the low performance. A simple SLURM job submission script would look like the following:

 #!/bin/bash -l
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=12
#SBATCH --ntasks-per-core=2
#SBATCH --cpus-per-task=2
#SBATCH --constraint=gpu
#SBATCH --time=00:30:00
export CRAY_CUDA_MPS=1
module load daint-gpu
srun -n $SLURM_NTASKS --ntasks-per-node=$SLURM_NTASKS_PER_NODE -c $SLURM_CPUS_PER_TASK ./test.exe

The flag -l at the beginning allows you to call the module command within the script, for instance to load the GPU enabled software stack in the MODULEPATH with the module daint-gpu. Alternatively, the module daint-mc will make the multicore software stack available in your MODULEPATH instead.

The SLURM option --constraint=gpu makes sure that the SLURM scheduler allocates the XC50 Intel Haswell 12-core nodes with GPU devices and automatically sets the option --gres=gpu:1.

The SLURM option --constraint=mc ensures that the batch job is allocated to the multicore XC40 Intel Broadwell 2 x 18-core nodes and not to the GPU nodes. You need to use this constraint when selecting the partition prepost.

It is required to specify either --constraint=gpu or --constraint=mc.  The srun flags -n, --ntasks-per-node and -c takes as argument the number of MPI tasks allocated, the number of MPI tasks per node and the number of threads per MPI task: please have a look at man srun for more details on these options.

You can target the large memory nodes on the multicore portion of the system using the SLURM flag --mem=120GB.

SLURM batch queues

Name of the queue Max time Max nodes Brief Description
debug 30min 4 Quick turnaround for test jobs
low 6 h 2400 For use only when allocations are exhausted
normal 24 h 2400 Standard queue for production work
high 24 h 2400 High priority queue, time is charged double
large 12 h 4400 Large scale work, by arrangement only
prepost 30min 1 High priority pre/post processing
xfer 24h 1 Data transfer queue
total 2 h CSCS maintenance queue (restricted use)

The list of queues and partitions is available typing sinfo or scontrol show partition. Note that not all groups are enabled on every partition, please check the AllowGroups entry of the command scontrol show partition.

You can choose the queue where to run your job by issuing the "--partition" directive in your batch script: #SBATCH --partition=<partition_name>

Please check the man pages and the official documentation for further details on SLURM directives.
For a list of the most useful SLURM commands, have a look at the corresponding entry in the FAQ list.
If you have any questions, don't hesitate to contact us at help(at)cscs.ch