Batch System Slurm¶

ZIH uses the batch system Slurm for resource management and job scheduling. Compute nodes are not accessed directly, but addressed through Slurm. You specify the needed resources (cores, memory, GPU, time, ...) and Slurm will schedule your job for execution.

When logging in to ZIH systems, you are placed on a login node. There, you can manage your data life cycle, setup experiments, and edit and prepare jobs. The login nodes are not suited for computational work! From the login nodes, you can interact with the batch system, e.g., submit and monitor your jobs.

Batch System

The batch system is the central organ of every HPC system users interact with its compute resources. The batch system finds an adequate compute system (partition) for your compute jobs. It organizes the queueing and messaging, if all resources are in use. If resources are available for your job, the batch system allocates and connects to these resources, transfers runtime environment, and starts the job.

A workflow could look like this:

sequenceDiagram
    user ->>+ login node: run programm
    login node ->> login node: kill after 5 min
    login node ->>- user: Killed!
    user ->> login node: salloc [...]
    login node ->> Slurm: Request resources
    Slurm ->> user: resources
    user ->>+ allocated resources: srun [options] [command]
    allocated resources ->> allocated resources: run command (on allocated nodes)
    allocated resources ->>- user: program finished
    user ->>+ allocated resources: srun [options] [further_command]
    allocated resources ->> allocated resources: run further command
    allocated resources ->>- user: program finished
    user ->>+ allocated resources: srun [options] [further_command]
    allocated resources ->> allocated resources: run further command
    Slurm ->> allocated resources: Job limit reached/exceeded
    allocated resources ->>- user: Job limit reached

Batch Job

At HPC systems, computational work and resource requirements are encapsulated into so-called jobs. In order to allow the batch system an efficient job placement it needs these specifications:

requirements: number of nodes and cores, memory per core, additional resources (GPU)
maximum run-time
HPC project for accounting
who gets an email on which occasion

Moreover, the runtime environment as well as the executable and certain command-line arguments have to be specified to run the computational work.

This page provides a brief overview on

Slurm options to specify resource requirements,
how to submit interactive and batch jobs,
how to write job files,
how to manage and control your jobs.

If you are are already familiar with Slurm, you might be more interested in our collection of job examples and job examples for GPU usage. There is also a ton of external resources regarding Slurm. We recommend these links for detailed information:

slurm.schedmd.com provides the official documentation comprising manual pages, tutorials, examples, etc.
Comparison with other batch systems

Job Submission¶

There are three basic Slurm commands for job submission and execution:

srun: Run a parallel application (and, if necessary, allocate resources first).
sbatch: Submit a batch script to Slurm for later execution.
salloc: Obtain a Slurm job allocation (i.e., resources like CPUs, nodes and GPUs) for interactive use. Release the allocation when finished.

Executing a program with srun directly on the shell will be blocking and launch an interactive job. Apart from short test runs, it is recommended to submit your jobs to Slurm for later execution by using batch jobs. For that, you can conveniently put the parameters in a job file, which you can submit using sbatch [options] <job file>.

After submission, your job gets a unique job ID, which is stored in the environment variable SLURM_JOB_ID at job runtime. The command sbatch outputs the job ID to stderr. Furthermore, you can find it via squeue --me. The job ID allows you to manage and control your jobs.

srun vs. mpirun

On ZIH systems, srun is used to run your parallel application. The use of mpirun is provenly broken on clusters Power9 and Alpha for jobs requiring more than one node. Especially when using code from github projects, double-check its configuration by looking for a line like 'submit command mpirun -n $ranks ./app' and replace it with 'srun ./app'.

Otherwise, this may lead to wrong resource distribution and thus job failure, or tremendous slowdowns of your application.

Options¶

The following table contains the most important options for srun, sbatch, salloc to specify resource requirements and control communication.

Options Table (see man sbatch for all available options)

Slurm Option	Description
`-n, --ntasks=<N>`	Total number of (MPI) tasks (default: 1)
`-N, --nodes=<N>`	Number of compute nodes
`--ntasks-per-node=<N>`	Number of tasks per allocated node to start (default: 1)
`-c, --cpus-per-task=<N>`	Number of CPUs per task; needed for multithreaded (e.g. OpenMP) jobs; typically `N` should be equal to `OMP_NUM_THREADS`
`--mem-per-cpu=<size>`	Memory need per allocated CPU in MB
`-t, --time=<HH:MM:SS>`	Maximum runtime of the job
`--mail-user=<your email>`	Get updates about the status of the jobs
`--mail-type=ALL`	For what type of events you want to get a mail; valid options: `ALL`, `BEGIN`, `END`, `FAIL`, `REQUEUE`
`-J, --job-name=<name>`	Name of the job shown in the queue and in mails (cut after 24 chars)
`--no-requeue`	Disable requeueing of the job in case of node failure (default: enabled)
`--exclusive`	Exclusive usage of compute nodes; you will be charged for all CPUs/cores on the node
`-A, --account=<project>`	Charge resources used by this job to the specified project
`-o, --output=<filename>`	File to save all normal output (stdout) (default: `slurm-%j.out`)
`-e, --error=<filename>`	File to save all error output (stderr) (default: `slurm-%j.out`)
`-a, --array=<arg>`	Submit an array job (examples)
`-w <node1>,<node2>,...`	Restrict job to run on specific nodes only
`-x <node1>,<node2>,...`	Exclude specific nodes from job
`--switches=<count>[@max-time]`	Optimum switches and max time to wait for optimum
`--signal=<sig_num>[@sig_time]`	Send signal `sig_num` to job `sig_time` before time limit (see Checkoint/Restart page)
`--test-only`	Retrieve estimated start time of a job considering the job queue; does not actually submit the job nor run the application

Output and Error Files

When redirecting stderr and stderr into a file using --output=<filename> and --stderr=<filename>, make sure the target path is writeable on the compute nodes, i.e., it may not point to a read-only mounted filesystem like /projects.

No free lunch

Runtime and memory limits are enforced. Please refer to the page Slurm resource limits for a detailed overview.

Host List¶

If you want to place your job onto specific nodes, use -w, --nodelist=<host1,host2,..> with a list of hosts that will work for you.

Number of Switches¶

You can fine tune your job by specifying the number of switches desired for the job allocation and optionally the maximum time to wait for that number of switches. The corresponding option to sbatch is --switches=<count>[@max-time]. The job remains pending until it either finds an allocation with desired switch count or the time limit expires. Acceptable time formats include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds". For a detailed explanation, please refer to the sbatch online documentation.

Interactive Jobs¶

Interactive activities like editing, compiling, preparing experiments etc. are normally limited to the login nodes. For longer interactive sessions, you can allocate resources on the compute node with the command salloc. It takes the same options as sbatch to specify the required resources.

salloc returns a new shell on the node where you submitted the job. You need to use the command srun in front of the following commands to have these commands executed on the allocated resources. If you request for more than one task, please be aware that srun will run the command on each allocated task by default! To release the allocated resources, invoke the command exit or scancel <jobid>.

Example: Interactive allocation using salloc

The following code listing depicts the allocation of two nodes with two tasks on each node with a time limit of one hour on the cluster Barnard for interactive usage.

marie@login.barnard$ salloc --nodes=2 --ntasks-per-node=2 --time=01:00:00
salloc: Pending job allocation 1234567
salloc: job 1234567 queued and waiting for resources
salloc: job 1234567 has been allocated resources
salloc: Granted job allocation 1234567
salloc: Waiting for resource configuration
salloc: Nodes n[1184,1223] are ready for job
[...]
marie@login.barnard$ hostname
login1
marie@login.barnard$ srun hostname
n1184
n1184
n1223
n1223
marie@login.barnard$ exit # ending the resource allocation

After Slurm successfully allocated resources for the job, a new shell is created on the submit host (cf. lines 9-10).

In order to use the allocated resources, you need to invoke your commands with srun (cf. lines 11 ff).

The command srun also creates an allocation, if it is running outside any sbatch or salloc allocation.

marie@login$ srun --pty --ntasks=1 --cpus-per-task=4 --time=1:00:00 --mem-per-cpu=1700 bash -l
srun: job 13598400 queued and waiting for resources
srun: job 13598400 has been allocated resources
marie@compute$ # Now, you can start interactive work with e.g. 4 cores

Since Slurm 20.11 --exclusive is the default for srun as a step, that means you have to use --overlap, if you want to run srun within a srun allocation.

marie@login$ srun --pty bash -l
srun: job 27410688 queued and waiting for resources
srun: job 27410688 has been allocated resources
marie@compute$ srun --overlap hostname
taurusi6604.taurus.hrsk.tu-dresden.de

Using module commands in interactive mode

The module commands are made available by sourcing the files /etc/profile and ~/.bashrc. This is done automatically by passing the parameter -l to your shell, as shown in the example above. If you missed adding -l at submitting the interactive session, no worry, you can source this files also later on manually (source /etc/profile).

Interactive X11/GUI Jobs¶

Slurm will forward your X11 credentials to the first (or even all) node for a job with the (undocumented) --x11 option.

marie@login$ srun --ntasks=1 --pty --x11=first xeyes

X11 error

If you are getting the error:

srun: error: x11: unable to connect node taurusiXXXX

that probably means you still have an old host key for the target node in your ~.ssh/known_hosts file (e.g. from pre-SCS5). This can be solved either by removing the entry from your known_hosts or by simply deleting the known_hosts file altogether if you don't have important other entries in it.

Batch Jobs¶

Working interactively using srun and salloc is a good starting point for testing and compiling. But, as soon as you leave the testing stage, we highly recommend to use batch jobs. Batch jobs are encapsulated within job files and submitted to the batch system using sbatch for later execution. A job file is basically a script holding the resource requirements, environment settings and the commands for executing the application. Using batch jobs and job files has multiple advantages*:

You can reproduce your experiments and work, because all steps are saved in a file.
You can easily share your settings and experimental setup with colleagues.

*) If job files are version controlled or environment env is saved along with Slurm output.

Syntax: Submitting a batch job

marie@login$ sbatch [options] <job_file>

Job Files¶

Job files have to be written with the following structure.

#!/bin/bash
# ^Batch script starts with shebang line

#SBATCH --ntasks=24                   # #SBATCH lines request resources and
#SBATCH --time=01:00:00               # specify Slurm options
#SBATCH --account=<KTR>               #
#SBATCH --job-name=fancyExp           # All #SBATCH lines have to follow uninterrupted
#SBATCH --output=simulation-%j.out    # after the shebang line
#SBATCH --error=simulation-%j.err     # Comments start with # and do not count as interruptions

module purge                          # Set up environment, e.g., clean/switch modules environment
module load <module1 module2>         # and load necessary modules

srun ./application [options]          # Execute parallel application with srun

The following two examples show the basic resource specifications for a pure OpenMP application and a pure MPI application, respectively. Within the section Job Examples, we provide a comprehensive collection of job examples.

Job file OpenMP

#!/bin/bash

#SBATCH --nodes=1
#SBATCH --tasks-per-node=1
#SBATCH --cpus-per-task=64
#SBATCH --time=01:00:00
#SBATCH --account=<account>

module purge
module load <modules>

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
srun ./path/to/openmp_application

Submisson: marie@login$ sbatch batch_script.sh
Run with fewer CPUs: marie@login$ sbatch --cpus-per-task=14 batch_script.sh

Job file MPI

#!/bin/bash

#SBATCH --ntasks=64
#SBATCH --time=01:00:00
#SBATCH --account=<account>

module purge
module load <modules>

srun ./path/to/mpi_application

Submisson: marie@login$ sbatch batch_script.sh
Run with fewer MPI tasks: marie@login$ sbatch --ntasks=14 batch_script.sh

Using Simultaneous Multithreading (SMT)¶

Most modern architectures offer simultaneous multithreading (SMT), where physical cores of a CPU are split into virtual cores (aka. threads). This technique allows to run two instruction streams per physical core in parallel.

At ZIH systems, SMT is available at the partitions rome and alpha. It is deactivated by default, because the environment variable SLURM_HINT is set to nomultithread. If you wish to make use of the SMT cores, you need to explicitly activate it. In principle, there are two different ways:

Change the value of the environment variable via export SLURM_HINT=multithread in your current shell and submit your job file, or invoke your srun or salloc command line.
Clear the environment variable via unset SLURM_HINT and provide the option --hint=multithread to sbatch, srun or salloc command line.

Warning

If you like to activate SMT via the directive

#SBATCH --hint=multithread

within your job file, you also have to clear the environment variable SLURM_HINT before submitting the job file. Otherwise, the environment varibale SLURM_HINT takes precedence.

Heterogeneous Jobs¶

A heterogeneous job consists of several job components, all of which can have individual job options. In particular, different components can use resources from different Slurm partitions. One example for this setting is an MPI application consisting of a master process with a huge memory footprint and worker processes requiring GPU support.

The salloc, sbatch and srun commands can all be used to submit heterogeneous jobs. Resource specifications for each component of the heterogeneous job should be separated with ":" character. Running a job step on a specific component is supported by the option --het-group.

marie@login$ salloc --ntasks=1 --cpus-per-task=4 --partition <partition> --mem=200G : \
                    --ntasks=8 --cpus-per-task=1 --gres=gpu:8 --mem=80G --partition <partition>
[...]
marie@login$ srun ./my_application <args for master tasks> : ./my_application <args for worker tasks>

Heterogeneous jobs can also be defined in job files. There, it is required to separate multiple components by a line containing the directive #SBATCH hetjob.

#!/bin/bash

#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --partition=<partition>
#SBATCH --mem=200G
#SBATCH hetjob # required to separate groups
#SBATCH --ntasks=8
#SBATCH --cpus-per-task=1
#SBATCH --gres=gpu:8
#SBATCH --mem=80G
#SBATCH --partition=<partition>

srun ./my_application <args for master tasks> : ./my_application <args for worker tasks>

# or as an alternative
srun ./my_application <args for master tasks> &
srun --het-group=1 ./my_application <args for worker tasks> &
wait

Limitations¶

Due to the way scheduling algorithm works it is required that each component has to be allocated on a different node. Furthermore, job arrays of heterogeneous jobs are not supported.

Manage and Control Jobs¶

Job and Slurm Monitoring¶

On the command line, use squeue to watch the scheduling queue.

Show your jobs

Invoke squeue --me to list only your jobs.

In its last column, the squeue command will also tell why a job is not running. Possible reasons and their detailed descriptions are listed in the following table. More information about job parameters can be obtained with scontrol -d show job <jobid>.

Reason Table

Reason	Long Description
`Dependency`	This job is waiting for a dependent job to complete.
`None`	No reason is set for this job.
`PartitionDown`	The partition required by this job is in a down state.
`PartitionNodeLimit`	The number of nodes required by this job is outside of its partitions current limits. Can also indicate that required nodes are down or drained.
`PartitionTimeLimit`	The jobs time limit exceeds its partitions current time limit.
`Priority`	One or higher priority jobs exist for this partition.
`Resources`	The job is waiting for resources to become available.
`NodeDown`	A node required by the job is down.
`BadConstraints`	The jobs constraints can not be satisfied.
`SystemFailure`	Failure of the Slurm system, a filesystem, the network, etc.
`JobLaunchFailure`	The job could not be launched. This may be due to a filesystem problem, invalid program name, etc.
`NonZeroExitCode`	The job terminated with a non-zero exit code.
`TimeLimit`	The job exhausted its time limit.
`InactiveLimit`	The job reached the system inactive limit.

For detailed information on why your submitted job has not started yet, you can use the command

marie@login$ whypending <jobid>

Editing Jobs¶

Jobs that have not yet started can be altered. By using scontrol update timelimit=4:00:00 jobid=<jobid>, it is for example possible to modify the maximum runtime. scontrol understands many different options, please take a look at the scontrol documentation for more details.

Canceling Jobs¶

The command scancel <jobid> kills a single job and removes it from the queue. By using scancel -u <username>, you can send a canceling signal to all of your jobs at once.

Evaluating Jobs¶

The Slurm command sacct provides job statistics like memory usage, CPU time, energy usage etc. as table-formatted output on the command line.

The job monitor PIKA provides web-based graphical performance statistics at no extra cost.

Learn from old jobs

We highly encourage you to inspect your previous jobs in order to better estimate the requirements, e.g., runtime, for future jobs. With PIKA, it is e.g. easy to check whether a job is hanging, idling, or making good use of the resources.

Using sacct (see also man sacct)

sacct outputs the following fields by default.

# show all own jobs contained in the accounting database
marie@login$ sacct
    JobID    JobName  Partition    Account  AllocCPUS      State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
[...]

We'd like to point your attention to the following options to gain insight in your jobs.

Show specific job

marie@login$ sacct --jobs=<JOBID>

Show all fields for a specific job

marie@login$ sacct --jobs=<JOBID> --format=All

Show specific fields

marie@login$ sacct --jobs=<JOBID> --format=JobName,MaxRSS,MaxVMSize,CPUTime,ConsumedEnergy

The manual page (man sacct) and the sacct online reference provide a comprehensive documentation regarding available fields and formats.

Time span

By default, sacct only shows data of the last day. If you want to look further into the past without specifying an explicit job id, you need to provide a start date via the option --starttime (or short: -S). A certain end date is also possible via --endtime (or -E).

Show all jobs since the beginning of year 2021

marie@login$ sacct --starttime 2021-01-01 [--endtime now]

Jobs at Reservations¶

Within a reservation, you have privileged access to HPC resources. How to ask for a reservation is described in the section reservations. After we agreed with your requirements, we will send you an e-mail with your reservation name. Then, you could see more information about your reservation with the following command:

marie@login$ scontrol show res=<reservation name>
# e.g. scontrol show res=hpcsupport_123

If you want to use your reservation, you have to add the parameter --reservation=<reservation name> either in your job script or to your srun or salloc command.

Node-Local Storage in Jobs¶

For some workloads and applications, it is valuable to use node-local storage in order to reduce or even completely omit usage of the parallel filesystems.

The availability and the capacity of local storage differ between the clusters, as depicted in the following table.

Cluster	Number of Nodes	Local Storage	Mountpoint	Request
`Alpha Centauri`	All compute nodes	3.5 TB on NVMe device	`/tmp`	Always present, no action needed
`Barnard`	12 nodes	1.8 TB on NVMe device	`/tmp`	`--constraint=local_disk` option to `sbatch`, `salloc`, and `srun`
`Capella`	All compute nodes	800 GB	`/tmp`	Always present, no action needed
`Romeo`	All compute nodes	200 GB	`/tmp`	Always present, no action needed

Clusters Power9 and Julia

Node-local storage is not available on the two clusters Power9 and Julia.

!!! hint Request nodes with local storage on Barnard

Note that most nodes on `Barnard` don't have a local disk and space in `/tmp` is **very**
limited. If you need a local disk request this with the
[Slurm feature](slurm.md#node-local-storage-in-jobs) `--constraint=local_disk` to
`sbatch`, `salloc`, and `srun`.