Introduction

At D-ITET the Slurm job scheduling sytem can be used for running compute-intensive jobs. It consists of a master host, where the scheduler resides and the compute nodes, where batch jobs are executed. The compute nodes are powerful servers located in server rooms, they are exclusively reserved for batch processing. Interactive logins are disabled.

Access

Access to the Slurm cluster is reserved for staff of the contributing institutes APS, MINS, TIK and PBL.
Access is granted on request:

If your circumstances differ and you'd still like to use the cluster, please contact ISG D-ITET support as well and ask for an offer. Time-limited test accounts for up to 2 weeks are also available on request.

Additional information for institutes

Some institutes have additional setup and configuration, if you are a member of such an institute, make sure to read the information linked below after reading this article:

Additional information for courses

D-ITET runs an additional GPU cluster reserved for official courses. Information specific to this cluster are available in this article:

Slurm

Slurm (Simple Linux Utility for Resource Management) is a free and open-source job scheduler for Linux and Unix-like kernels, used by many of the world's supercomputers and compute clusters. Slurm's design is very modular with about 100 optional plugins. In 2010, the developers of Slurm founded SchedMD, which maintains the canonical source, provides development, level 3 commercial support and training services and also provide a very good online documentation to Slurm.

Slurm Cluster

Hardware

At the moment the computing power of the Slurm cluster is based on the following 11 cpu compute nodes and 1 gpu compute node:

Server

CPU

Frequency

Cores

Memory

/scratch SSD

/scratch Size

GPUs

GPU Memory

Operating System

arton[01-03]

Dual Octa-Core Intel Xeon E5-2690

2.90 GHz

16

125 GB

-

895 GB

-

-

Debian 10

arton[04-08]

Dual Deca-Core Intel Xeon E5-2690 v2

3.00 GHz

20

125 GB

-

895 GB

-

-

Debian 10

arton[09-10]

Dual Deca-Core Intel Xeon E5-2690 v2

3.00 GHz

20

251 GB

1.7 TB

-

-

Debian 10

arton11

Dual Deca-Core Intel Xeon E5-2690 v2

3.00 GHz

20

535 GB

1.7 TB

-

-

Debian 10

artongpu01

Dual Octa-Core Intel Xeon Silver 4208

2.10 GHz

16

125 GB

1.1 TB

4 RTX 2080 Ti

11 GB

Debian 10

The nodes are "weighted", which gives the scheduler an additional selection criteria between nodes which fulfill criterias to run a job, like resources and membership in certain partitions. The idea is to prefer nodes with faster CPUs and of those, prefer those with lower RAM. For details, see /home/sladmitet/slurm/nodes.conf.

The Slurm job scheduler runs on the linux server itetmaster01.

Software

The nodes offer the same software environment as all D-ITET managed Linux clients, gpu nodes have a restricted software (no desktops installed, minimal dependencies needed for driver support).

Using Slurm

At a basic level, Slurm is very easy to use. The following sections will describe the commands you need to run and manage your batch jobs. The commands that will be most useful to you are as follows:

Setting environment

The above commands only work if the environment variables for Slurm are set. Please issue the following command in your bash shell to start working with the cluster immediately or add them to your ~/.bashrc to reference the Slurm cluster for new instances of bash:

export SLURM_CONF=/home/sladmitet/slurm/slurm.conf

sbatch → Submitting a job

sbatch doesn't allow to submit a binary program directly, wrap the program to run into a surrounding bash script. The sbatch command has the following syntax:

> sbatch [temporary_options] job_script [job_script arguments]

The job_script is a standard UNIX shell script. The fixed options for the Slurm Scheduler are placed in the job_script in lines starting with #SBATCH. The UNIX shell interprets these lines as comments and ignores them.

To test your job-script simply run it interactively on your host.

Assume there is a c program primes.c which is compiled to an executable binary with "gcc -o primes primes.c" and stored as /absolute/path/to/primes. The program runs 5 seconds and calculates prime numbers. The found prime numbers and a final summary are sent to standard output. A sample job_script placed in the same location /absolute/path/to/primes.sh to perform a batch run of the binary primes on the Arton cluster looks like this:

#!/bin/bash

#SBATCH --mail-type=ALL                           # mail configuration: NONE, BEGIN, END, FAIL, REQUEUE, ALL
#SBATCH --output=/absolute/path/to/log/%j.out     # where to store the output (%j is the JOBID), subdirectory "log" must exist
#SBATCH --error=/absolute/path/to/log/log/%j.err  # where to store error messages

# Exit on errors
set -o errexit

# Set a directory for temporary files unique to the job with automatic removal at job termination
TMPDIR=$(mktemp -d)
if [[ ! -d ${TMPDIR} ]]; then
    echo 'Failed to create temp directory' >&2
    exit 1
fi
trap "exit 1" HUP INT TERM
trap 'rm -rf "${TMPDIR}"' EXIT
export TMPDIR

# Change the current directory to the location where you want to store temporary files, exit if changing didn't succeed.
# Adapt this to your personal preference
cd "${TMPDIR}" || exit 1

# Send some noteworthy information to the output log
echo "Running on node: $(hostname)"
echo "In directory:    $(pwd)"
echo "Starting on:     $(date)"
echo "SLURM_JOB_ID:    ${SLURM_JOB_ID}"

# Binary or script to execute
/absolute/path/to/primes

# Send more noteworthy information to the output log
echo "Finished at:     $(date)"

# End the script with exit code 0
exit 0

You can test the script by running it interactively in a terminal:

$ /absolute/path/to/primes.sh

If the script runs successfully you can now submit it as a batch job to the Slurm arton cluster:

$ sbatch /absolute/path/to/primes.sh
sbatch: Start executing function slurm_job_submit......
sbatch: Job partition set to : cpu.normal
Submitted batch job 931

After the job has finished, you will find the output file of the job in the file /absolute/path/to/log/<JOBID>.out. If there were errors they are stored in the file /absolute/path/to/log/<JOBID>.err.
⚠ Remember: The directory for the job output has to exist before submitting the job, it is not created automatically!

You can only submit jobs to Slurm if your account is configured in the Slurm user database. If it isn't, you'll receive this error message

sbatch → Submitting an array job

It is also possible to start an array job. The above job would run 10 times if you added the option #SBATCH --array=0-9 to the job-script. A repeated execution only makes sense if the executed program adapts its behaviour according to the changing array task count number. The array count number can be referenced through the variable $SLURM_ARRAY_TASK_ID. You can pass the value of $SLURM_ARRAY_TASK_ID or some derived parameters to the executable.
Here is a simple example of passing an input filename parameter changing with $SLURM_ARRAY_TASK_ID to the executable:

.
#SBATCH   --array=0-9
#
# binary to execute
<path-to-executable> data$SLURM_ARRAY_TASK_ID.dat

Every run of the program in the array job with a different task-id will produce a separate output file.

The option expects a range of task-ids expressed in the form --array=n[,k[,...]][-m[:s]]%l
where n, k, m are discreet task-ids, s is a step applied to a range n-m and l applies a limit to the number of simultaneously running tasks. See man sbatch for examples.
Specifying one task-id instead of a range as in --array=10 results in an array job with a single task with task-id 10.
The following variables will be available in the job context and reflect the option arguments given: $SLURM_ARRAY_TASK_MAX, $SLURM_ARRAY_TASK_MIN, $SLURM_ARRAY_TASK_STEP.

sbatch → Common options

The following table shows the most common options available for sbatch to be used in the job_script in lines starting with #SBATCH

option

description

--mail-type=...

Possible Values: NONE, BEGIN, END, FAIL, REQUEUE, ALL

--mem=<n>G

the job needs a maximum of <n> GByte ( if omitted the default of 6G is used )

--cpus-per-task=<n>

number of cores to be used for the job

--gres=gpu:1

number of GPUs needed for the job

--nodes=<n>

number of compute nodes to be used for the job

--hint=<type>

Bind tasks to CPU cores according to application hints (See man --pager='less +/--hint' srun and multi-core support

--constraint=<feature_name>

Request one or more features, optionally combined by operators (See man --pager='less +/--constraint' sbatch)

squeue → Show running/waiting jobs

The squeue command shows the actual list of running and pending jobs in the system. As you can see in the following sample output the default format is quite minimalistic:

$ squeue
JOBID PARTITION     NAME     USER ST  TIME  NODES  NODELIST(REASON)
  951 cpu.norma primes.s gfreudig  R  0:11      1  arton02
  950 cpu.norma primes_4 gfreudig  R  0:36      1  arton02
  949 cpu.norma primes.s fgtest01  R  1:22      1  arton02
  948 gpu.norma primes.s fgtest01  R  1:39      1  artongpu01

More detailed information can be obtained by issuing the following command:

$ squeue --Format=jobarrayid:10,state:10,partition:16,reasonlist:18,username:10,tres-alloc:45,timeused:8,command:50
JOBID  STATE    PARTITION   NODELIST(REASON)  USER      TRES_ALLOC                                TIME  COMMAND                                             
951    RUNNING  cpu.normal  arton02           gfreudig  cpu=1,mem=32G,node=1,billing=1            1:20  /home/gfreudig/BTCH/Slurm/jobs/single/primes.sh 600 
950    RUNNING  cpu.normal  arton02           gfreudig  cpu=4,mem=8G,node=1,billing=4             1:45  /home/gfreudig/BTCH/Slurm/jobs/multi/primes_4.sh 600
949    RUNNING  cpu.normal  arton02           fgtest01  cpu=1,mem=8G,node=1,billing=1             2:31  /home/fgtest01/BTCH/Slurm/jobs/single/primes.sh 600 
948    RUNNING  gpu.normal  artongpu01        fgtest01  cpu=1,mem=8G,node=1,billing=1,gres/gpu=1  2:48  /home/fgtest01/BTCH/Slurm/jobs/single/primes.sh 600 

Defining an alias in your .bashrc with

alias sq1='squeue --Format=jobarrayid:10,state:10,partition:16,reasonlist:18,username:10,tres-alloc:45,timeused:8,command:50'

puts the command sq1 at your fingertips.

Never call squeue from any kind of loop, i.e. never do watch squeue. See man --pager='less +/^PERFORMANCE' squeue for an explanation.
To monitor your jobs, set the sbatch option --mail-type to send you notifications. If you absolutely have to see a live display of your jobs, use the --iterate option with a value of several seconds:

squeue --user=$USER --iterate=30

squeue → Show job steps

Individual job steps are listed with a specific option:

squeue -s

scancel → Deleting a job

With scancel you can remove your waiting and running jobs from the scheduler queue by their associated JOBID. The command squeue lists your jobs including their JOBIDs. A job can then be deleted with

> scancel <JOBID>

To operate on an array job you can use the following commands

> scancel <JOBID>          # all jobs (waiting or running) of the array job are deleted
> scancel <JOBID>_n        # the job with task-ID n is deleted
> scancel <JOBID>_[n1-n2]  # the jobs with task-ID in the range n1-n2 are deleted

sinfo → Show partition configuration

The partition status can be obtained by using the sinfo command. An example listing is shown below.

PARTITION      AVAIL  TIMELIMIT  NODES  STATE NODELIST
cpu.normal*    up 7-00:00:00     10   idle arton[01-11]
gpu.normal     up 2-00:00:00      1   idle artongpu01
tikgpu.all     up 2-00:00:00      7   idle tikgpu[01-07]
tikgpu.medium  up 2-00:00:00      3   idle tikgpu[01-03]

The partition is chosen by the scheduler according to your resource request and memberships in Slurm accounts. The logic can be seen in /home/sladmitet/slurm/jobsumit.lua.

sinfo → Show resources and utilization

Adding selected format parameters to the sinfo command shows the resources available on every node and their utilization:

sinfo -Node --Format nodelist:12,statecompact:7,memory:7,allocmem:10,freemem:10,cpusstate:15,cpusload:10,gresused:100 |(sed -u 1q; sort -u)

Restricting the command to a selected partition allows to show only GPU nodes:

sinfo -Node --partition=gpu.normal --Format nodelist:12,statecompact:7,memory:7,allocmem:10,freemem:10,cpusstate:15,cpusload:10,gresused:100

sinfo → Show available features

So-called features are used to constrain jobs to nodes with different hardware capabilities, typically GPU types. To show currently active features issue the following command sequence:

sinfo --Format nodehost:20,features_act:80 |grep -v '(null)' |awk 'NR == 1; NR > 1 {print $0 | "sort -n"}'

An example of feature use can be seen in section Specifying GPUs based on compute capability.

srun → Start an interactive shell

An interactive session on a compute node is possible for short tests, checking your environment or transferring data to the local scratch of a node available under /scratch_net/arton[0-11]. Such sessions are limited to a maximum run time of 720 minutes (12 hours) regardless of the partition they are sent to.

An interactive session lasting for 10 minutes on a GPU node can be started with:

srun --time 10 --gres=gpu:1 --pty bash -i

The ouptut will look similar to the following:

srun: Start executing function slurm_job_submit......
srun: Your job is a gpu job.
srun: Setting partition to gpu.normal
srun: job 11526 queued and waiting for resources

Omitting the parameter --gres=gpu:1 opens an interactive session on a CPU-only node.
Do not use an interactive login to run compute jobs, use this only briefly as outlined above. Restrict job time to the necessary minimum with the --time option as shown above. For details see the related section in the srun man page by issuing the command man --pager='less +/--time' srun in your shell.

srun → Attaching an interactive shell to a running job

An interactive shell can be opened inside a running job by specifying its job id:

srun --time 10 --jobid=123456 --overlap --pty bash -i

A typical use case of the above is interactive live-monitoring of a running job.

srun → Launch a command as a job step

When srun is used inside a sbatch script it spawns the given command inside a job step. This allows resource monitoring with the sstat command (see man sstat. Spawning several single-threaded commands and putting them in the background allows to schedule these commands inside the job allocation.
Here's an example how to run overall GPU logging and per-process logging in job steps before starting the actual computing commands.

...
set -o errexit

srun --ntasks=1 --cpus-per-task=1 nvidia-smi dmon -i ${CUDA_VISIBLE_DEVICES} -d 5 -s ucm -o DT > "${SLURM_JOB_ID}.gpulog" &
srun --ntasks=1 --cpus-per-task=1 nvidia-smi pmon -i ${CUDA_VISIBLE_DEVICES} -d 5 -s um -o DT  > "${SLURM_JOB_ID}.processlog" &
...
echo finished at: `date`
exit 0;

sstat → Display status information of a running job

The status information shows your job's resource usage while it is running:

sstat --jobs=<JOBID> --allsteps --format=JobID,AveVMSize%15,MaxRSS%15,AveCPU%15

All your currently running job's resource usages can be shown with:

sstat --jobs=$(squeue --noheader --me --format=%A |paste -s -d ',') --allsteps --format=JobID,AveVMSize%15,MaxRSS%15,AveCPU%15

sacct → Display accounting information of past jobs

Accounting information for past jobs can be displayed with various details (see man page).
The following example lists all jobs of the logged in user since the beginning of the year 2020:

sacct --user ${USER} --starttime=2020-01-01 --format=JobID,Start%20,Partition%20,ReqTRES%50,AveVMSize%15,MaxRSS%15,AveCPU%15,Elapsed%15,State%20

sprio → Show priorities of pending jobs

Pending jobs are prioritized by the scheduler by accounting for fair sharing of resources and age of a pending job. Priorities of pending jobs and the factors comprising them can be shown with

sprio --long

On job submission, a nice value can be added to influence priorities of your own jobs:

sbatch --nice=10 job_script.sh

The nice value of an already pending job can be incremented with positive values:

scontrol update job <jobid> nice=5

Only incrementation is possible. The value can be reset to nice=0.

The offical Slurm manual contains a detailled explanation of job priorisation.

smon → GPU / CPU availability

Information about the GPU nodes and current availability of the installed GPUs as well as CPU availability of CPU-only nodes is updated every 5 minutes to the file /home/sladmitet/smon.txt. Here are some convenient aliases to display the file with highlighting of either free GPUs or those running the current user's jobs:

alias smon_free="grep --color=always --extended-regexp 'free|$' /home/sladmitet/smon.txt"
alias smon_mine="grep --color=always --extended-regexp '${USER}|$' /home/sladmitet/smon.txt"

For monitoring its content the following aliases can be used:

alias watch_smon_free="watch --interval 300 --no-title --differences --color \"grep --color=always --extended-regexp 'free|$' /home/sladmitet/smon.txt\""
alias watch_smon_mine="watch --interval 300 --no-title --differences --color \"grep --color=always --extended-regexp '${USER}|$' /home/sladmitet/smon.txt\""

⚠ Never use watch directly on smon, as this places considerable load on the Slurm controller!

GPU jobs

Selecting the allocated GPUs

To select the GPU allocated by the scheduler, Slurm sets the environment variable CUDA_VISIBLE_DEVICES in the context of a job to the GPUs allocated to the job. The numbering always starts at 0 and is consecutively numbered up to the requested amount of GPUs - 1.
It is imperative to work with this variable exactly as it is set by Slurm, anything else leads to unexpected errors.
For details see the section GPU Management in the official Slurm documentation.

Specifying a GPU type

It's possible to specify a GPU type by inserting the type description in the gres allocation:

--gres=gpu:titan_rtx:1

Available GPU type descriptions can be filtered from an appropriate sinfo command:

sinfo --noheader --Format gres:200 |tr ':' '\n' |sort -u |grep -vE '^(gpu|[0-9,\(]+)'

Multiple GPU types can be requested by using the --constraint option.

Specifying GPUs based on compute capability

CUDA code compiled with nvcc can be optimized for ranges of so-called Compute Capabilities defining generation.version of a NVIDIA GPU.
More information about compute capabilites can be read here:

The following table shows an abbreviated list of the compute capabilities of available GPU types selectable by features:

Compute capability

Features

3.5

tesla_k40c

5.2

geforce_gtx_titan_x

6.1

geforce_gtx_1080_ti, titan_x, titan_xp

7.0

tesla_v100

7.5

geforce_rtx_2080_ti, titan_rtx

8.0

a100

8.6

geforce_rtx_3090, a6000

For more information see the full list of NVIDIA CUDA GPUs.

As GPU nodes may house different generations of GPUs, compiled CUDA code might not run on all of them and errors similar to the following can appear:

RuntimeError: CUDA error: no kernel image is available for execution on the device

If you see a similar error:

Example: Add the constraint --constraint='tesla_v100|geforce_rtx_2080_ti|titan_rtx|geforce_rtx_3090' to your job submission to run a job only on nodes with GPUs of compute capability 7.0 or higher.

Multicore jobs/ job to core binding

A modern linux kernel is able to bind a process and all its children to a fixed number of cores. By default a job submitted to the Slurm arton cluster is bound to to the numbers of requested cores/cpus. The default number of requested cpus is 1, if you have an application which is able to run multithreaded on several cores you must use the --cpus-per-task option in the sbatch command to get a binding to more than one core. To check for processes with core bindings, use the command hwloc-ps -c:

$ ssh arton02 hwloc-ps -c
43369   0x00010001              slurmstepd: [984.batch]
43374   0x00010001              /bin/sh
43385   0x00010001              codebin/primes

Job input/output data storage

Temporary data storage of a job used only while the job is running, should be placed in the /scratch directory of the compute nodes. Set the environment variables of the tools you use accordingly. The Matlab MCR_ROOT_CACHE variable is set automatically by the Slurm scheduler.
The file system protection of the /scratch directory allows everybody to create files and directories in it. A cron job runs periodically on the execution hosts to prevent the /scratch directory from filling up and cleans it governed by pre-set policies. Therefore data you place in the /scratch directory of a compute node cannot be assumed to stay there forever.

Other data storage concepts for the arton cluster are possible and will be investigated, if the above solution proves not to be sufficient.

Matlab Distributed Computing Environment (MDCE)

The Matlab Parallel Computing Toolbox (PCT) can be configured with an interface to the Slurm cluster. To work with MDCE please import Slurm.mlsettings in Matlab GUI (Parallel → Create and manage Clusters → Import ). Adjust the setting JobStorageLocation to your requirements. The cluster profile Slurm will now appear besides the standard local(default) profile in the profile list. With the local profile, you can use as many workers on one computer as there are physical cores while the Slurm profile allows to initiate up to 32 worker processes distributed over all Slurm compute nodes.
⚠ Please temporay reduce the number of workers to 4 in the Slurm profile when performing the profile "Validation" function in the Matlab Cluster Manager.
⚠ Don't forget to set the Slurm environment variables before starting Matlab!
The Slurm cluster profile can be used with Matlab programs running as Slurm batch jobs but it's also possible to use the profile in an interactive Matlab session on your client. When you open a Slurm parpool, the workers are started automatically as jobs in the cluster.
⚠ In interactive mode please always close your parpool if you aren't performing any calculations on the workers.
Sample code for the 3 Matlab PCT methods parfor, spmd, tasks using the local or Slurm cluster profile is provided in PCTRefJobs.tar.gz.

Reservations

Nodes may be reserved at certain times for courses or maintenance. If your job is pending with the reason ReqNodeNotAvail,_May_be_reserved_for_other_job, check reservations and adjust the --time parameter of your job accordingly.

Showing current reservations

Current reservations can be shown by issuing

scontrol show reservation

Using a reservation

If you are entitled to use a reservation, specify the reservation in your job submission by appending the parameter --reservation=<ReservationName>.

Requesting a reservation

Reservations are managed by Slurm administrators. Please contact ISG D-ITET support if you're in need of a reservation.

Frequently Asked Questions

If your question isn't listed below, an answer might be listed in the official Slurm FAQ.

Batch job submission failed: Invalid account

If you receive one of the following error messages after submitting a job with sbatch or using srun

sbatch: error: Batch job submission failed: Invalid account or account/partition combination specified

srun: error: Unable to allocate resources: Invalid account or account/partition combination specified

your account hasn't been registered with Slurm yet. Please contact support and ask to be registered.

Code runs with srun but fails with sbatch

This points to a difference in the environment of the interactive job and the batch job.

By default both srun and sbatch forward the complete current submission environment to a job (See man --pager='less +/--export=' srun, man --pager='less +/--export=' sbatch) and also change path to the current path of the submission environment (See man --pager='less +/--chdir=' srun, man --pager='less +/--chdir=' sbatch)

Compare the output of printenv from jobs started with srun and sbatch to figure out the differences:

Differences have to come from:

Invalid user for SlurmUser slurm

After executing one of the Slurm executables like sbatch or sinfo the following error appears:

error: Invalid user for SlurmUser slurm, ignored

The user slurm doesn't exist on the host you're running your Slurm executable. If this happens on a host managed by ISG D-ITET, please contact support, tell us the name of your host and ask us to configure it as a Slurm submission host.

Node(s) in drain state

If sinfo shows one or more nodes in drain state, the reason can be shown with

sinfo -R

or in case the reason is cut off with

sinfo -o '%60E %9u %19H %N'

Nodes are set to drain by ISG D-ITET to empty them of jobs in time for scheduled maintenance or by the scheduler itself in case a problem is detected on a node.

<Slurm command>: fatal: Could not establish a configuration source

If you receive the following error messages after using a Slurm command like srun, sbatch, squeue or sinfo (replace squeue with the name of the command you used to end up with the error message):

squeue: fatal: Could not establish a configuration source

Make sure you set your environment.

My job was terminated by the OOM killer

If your job got terminated and you see a line similar to the following in your job log:

slurmstepd: error: Detected 1 oom-kill event(s) in StepId=<JOB_ID>.batch cgroup. ...

this means a process in your job attempted to use more memory than you requested for the job, so it was killed by the OOM (Out Of Memory) killer. This in turn resulted in termination of your job by the Slurm scheduler.
Check the value of MaxRSS in the output of sacct for your job to verify the maximum memory usage of your job. Run tests by adjusting your memory allocation with --mem until you figure out how much memory your job needs.
Slurm's accounting samples jobs every 30 seconds, so there is no useful data if a job was killed within the first 30 seconds after it started. Also sudden spikes in memory consumption may not be recorded, but can still trigger the OOM killer.
⚠ This is error pertains to the onboard memory allocated to a job, a GPU allocation always contains its full GPU memory.

Services/SLURM (last edited 2024-02-01 11:35:02 by stroth)