Scheduler Overview (Slurm): Difference between revisions

From HPCwiki
Jump to navigation Jump to search
← Older edit
Haars0011 (talk | contribs)
Bureaucrats, Administrators
189 edits
VisualWikitext
Phase 1 § 4 P1.4.1: trim to overview only — content split into Partitions / Queues, Choosing a node (constraints), Batch Jobs, Interactive Jobs, Cancelling Jobs, Monitoring Jobs (separate pages). This page is now the entry point with QoS overview and topic index. (via update-page on MediaWiki MCP Server)
 
(17 intermediate revisions by 3 users not shown)
Line 1: Line 1:
The resource allocation / scheduling software on Anunna is [http://en.wikipedia.org/wiki/Simple_Linux_Utility_for_Resource_Management SLURM]: '''S'''imple '''L'''inux '''U'''tility for '''R'''esource '''M'''anagement.
The resource allocation and scheduling software on Anunna is [http://en.wikipedia.org/wiki/Simple_Linux_Utility_for_Resource_Management SLURM]: '''S'''imple '''L'''inux '''U'''tility for '''R'''esource '''M'''anagement. This page is the entry point — most topics have their own page; below is a short summary plus links.


== What's on which page ==


== Queues and defaults ==
* [[Partitions / Queues]] — list of partitions (<code>main</code>, <code>gpu</code>, <code>gpu_amd</code>) and how to choose one.
* [[Choosing a node (constraints)]] — defaults, hardware constraints, GPU selection.
* [[Batch Jobs]] — writing sbatch scripts and submitting them, including multi-job submissions and dependencies.
* [[Interactive Jobs]] — <code>sinteractive</code> and <code>salloc</code> for live shell sessions on a compute node.
* [[Array jobs]] — running the same script many times with a varying parameter.
* [[Monitoring Jobs]] — <code>squeue</code>, <code>scontrol</code>, <code>sstat</code>, <code>sacct</code>, <code>node_usage_graph</code>.
* [[Cancelling Jobs]] — <code>scancel</code>.
* [[Reservations]] — booking nodes in advance for events.


=== Queues ===
== Quality of Service ==
Every organization has 3 queues (in slurm called partitions) : a high, a standard and a low priority queue.<br>
The High queue provides the highest priority to jobs (20) then the standard queue (10). In the low priority queue (0)<br>
jobs will be resubmitted if a job with higer priority needs cluster resources and those resoruces are occupied by a Low queue jobs.
To find out which queues your account has been authorized for, type sinfo:
<source lang='bash'>
PARTITION      AVAIL  TIMELIMIT  NODES  STATE NODELIST
ABGC_High      up  infinite    12  down* node[043-048,055-060]
ABGC_High      up  infinite      6    mix fat[001-002],node[002-005]
ABGC_High      up  infinite    44  idle node[001,006-042,049-054]
ABGC_Std      up  infinite    12  down* node[043-048,055-060]
ABGC_Std      up  infinite      6    mix fat[001-002],node[002-005]
ABGC_Std      up  infinite    44  idle node[001,006-042,049-054]
ABGC_Low      up  infinite    12  down* node[043-048,055-060]
ABGC_Low      up  infinite      6    mix fat[001-002],node[002-005]
ABGC_Low      up  infinite    44  idle node[001,006-042,049-054]
</source>


=== Defaults ===
When submitting a job, you may optionally assign a different Quality of Service (QoS) to it:
There is no default queue, so you need to specify which queue to use when submitting a job.<br>
'''The default run time for a job is 1 hour!''' <br>
'''Default memory limit is 100MB per node!'''


== Submitting jobs: sbatch ==
<syntaxhighlight lang="bash">
#SBATCH --qos=std
</syntaxhighlight>


=== Example ===
The QoS values configured on Anunna:
Consider this simple python3 script that should calculate Pi to 1 million digits:
<source lang='python'>
from decimal import *
D=Decimal
getcontext().prec=10000000
p=sum(D(1)/16**k*(D(4)/(8*k+1)-D(2)/(8*k+4)-D(1)/(8*k+5)-D(1)/(8*k+6))for k in range(411))
print(str(p)[:10000002])
</source>


=== Loading modules ===
* '''std''' (priority 10) — the default. Use this unless you have a specific reason to pick another.
In order for this script to run, the first thing that is needed is that Python3, which is not the default Python version on the cluster, is load into your environment. Availability of (different versions of) software can be checked by the following command:
* '''low''' (priority 1) — reduced priority, but limited to 8 hours per job so a flood of low-priority jobs cannot lock up the cluster.
  module avail
* '''high''' (priority 20) — higher priority than <code>std</code>. More expensive — see [[Tariffs]].
* '''interactive''' (priority 100) — the highest priority, exclusively for immediate-running interactive jobs. You may not submit many or large jobs at this QoS.


In the list you should note that python3 is indeed available to be loaded, which then can be loaded with the following command:
Jobs can in principle be restarted and rescheduled if a higher-priority job needs cluster resources, but at the time of writing this preemption is not actually configured.
  module load python/3.3.3


=== Batch script ===
== Running MPI jobs ==
[[Creating_sbatch_script | Main Article: Creating a sbatch script]]


The following shell/slurm script can then be used to schedule the job using the sbatch command:
For multi-node MPI workloads see [[MPI on B4F cluster | MPI on Anunna]].
<source lang='bash'>
#!/bin/bash
#SBATCH --comment=773320000
#SBATCH --time=1200
#SBATCH --mem=2048
#SBATCH --ntasks=1
#SBATCH --output=output_%j.txt
#SBATCH --error=error_output_%j.txt
#SBATCH --job-name=calc_pi.py
#SBATCH --partition=ABGC_Std
#SBATCH --mail-type=ALL
#SBATCH --mail-user=email@org.nl


== See also ==


time python3 calc_pi.py
* [[Partitions / Queues]]
</source>
* [[Choosing a node (constraints)]]
 
* [[Batch Jobs]]
=== Submitting ===
* [[Interactive Jobs]]
The script, assuming it was named 'run_calc_pi.sh', can then be posted using the following command:
* [[Array jobs]]
<source lang='bash'>
* [[Monitoring Jobs]]
sbatch run_calc_pi.sh
* [[Cancelling Jobs]]
</source>
* [[Reservations]]
 
=== Submitting multiple jobs (simple) ===
Assuming there are 10 job scripts, name runscript_1.sh through runscript_10.sh, all these scripts can be submitted using the following line of shell code:
<source lang='bash'>for i in `seq 1 10`; do echo $i; sbatch runscript_$i.sh;done
</source>
 
=== Submitting multiple jobs (complex) ===
Lets's say you have three job scripts that depend on each other:
 
<source lang='bash'>job_1.sh #A simple initialisation script</source>
<source lang='bash'>job_2.sh #An array task</source>
<source lang='bash'>job_3.sh #Some finishing script, single run, after everything previous has finished</source>
 
You can create a script to simultaneously submit each job with a dependency on each other:
 
<source lang='bash'>#!/bin/bash
JOB1=$(sbatch job_1.sh| rev | cut -d ' ' -f 1 | rev) #Get me the last space-separated element
 
if ! [ "z$JOB1" == "z" ] ; then
  echo "First job submitted as jobid $JOB1"
  JOB2=$(sbatch --dependency=afterany:$JOB1 job_2.sh| rev | cut -d ' ' -f 1 | rev)
 
  if ! [ "z$JOB2" == "z" ] ; then
  echo "Second job submitted as jobid $JOB2, following $JOB1"
  JOB3=$(sbatch --dependency=afterany:$JOB2 job_3.sh| rev | cut -d ' ' -f 1 | rev)
 
  if ! [ "z$JOB3" == "z" ] ; then
  echo "Third job submitted as jobid $JOB3, following after every element of $JOB2"
 
  fi
fi
fi
</source>
 
This will ensure that the subsequent jobs occur after any finishing of the former (even if they failed).
 
Please see [https://slurm.schedmd.com/sbatch.html#OPT_dependency the sbatch documentation] for other options available to you. Note that aftercorr makes a subsequent array jobs array elements start after the correspondingly numbered ones from the previous job.
 
=== Submitting array jobs ===
<source lang='bash'>
#SBATCH --array=0-10%4
</source>
SLURM allows you to submit multiple jobs using the same template. Further information about this can be found [[Array_jobs|here]].
 
=== Using /tmp ===
There is a local disk of ~300G that can be used to temporarily stage some of your workload attached to each node. This is free to use, but please remember to clean up your data after usage.
 
In order to be sure that you're able to use space in /tmp, you can add
<source lang='bash'>
#SBATCH --tmp=<required size>
</source>
To your sbatch script. This will prevent your job from being run on nodes where there is no free space, or it's aimed to be used by another job at the same time.
 
=== Using GPU ===
There are two GPU nodes, in order to run a job that uses GPU on one of these nodes, you can add
 
<source lang='bash'>
#SBATCH --reservation='GPU Testing'
</source>
To your sbatch script. Without this parameter, your job won't run on one of these nodes.
 
== Monitoring submitted jobs ==
Once a job is submitted, the status can be monitored using the <code>squeue</code> command. The <code>squeue</code> command has a number of parameters for monitoring specific properties of the jobs such as time limit.
 
=== Generic monitoring of all running jobs ===
<source lang='bash'>
  squeue
</source>
 
You should then get a list of jobs that are running at that time on the cluster, for the example on how to submit using the 'sbatch' command, it may look like so:
    JOBID PARTITION    NAME    USER  ST      TIME  NODES NODELIST(REASON)
  3396      ABGC BOV-WUR- megen002  R      27:26      1 node004
  3397      ABGC BOV-WUR- megen002  R      27:26      1 node005
  3398      ABGC BOV-WUR- megen002  R      27:26      1 node006
  3399      ABGC BOV-WUR- megen002  R      27:26      1 node007
  3400      ABGC BOV-WUR- megen002  R      27:26      1 node008
  3401      ABGC BOV-WUR- megen002  R      27:26      1 node009
  3385  research BOV-WUR- megen002  R      44:38      1 node049
  3386  research BOV-WUR- megen002  R      44:38      1 node050
  3387  research BOV-WUR- megen002  R      44:38      1 node051
  3388  research BOV-WUR- megen002  R      44:38      1 node052
  3389  research BOV-WUR- megen002  R      44:38      1 node053
  3390  research BOV-WUR- megen002  R      44:38      1 node054
  3391  research BOV-WUR- megen002  R      44:38      3 node[049-051]
  3392  research BOV-WUR- megen002  R      44:38      3 node[052-054]
  3393  research BOV-WUR- megen002  R      44:38      1 node001
  3394  research BOV-WUR- megen002  R      44:38      1 node002
  3395  research BOV-WUR- megen002  R      44:38      1 node003
 
=== Monitoring time limit set for a specific job ===
The default time limit is set at one hour. Estimated run times need to be specified when running jobs. To see what the time limit is that is set for a certain job, this can be done using the <code>squeue</code> command.
<source lang='bash'>
squeue -l -j 3532
</source>
Information similar to the following should appear:
  Fri Nov 29 15:41:00 2013
  JOBID PARTITION    NAME    USER    STATE      TIME TIMELIMIT  NODES NODELIST(REASON)
  3532      ABGC BOV-WUR- megen002  RUNNING    2:47:03 3-08:00:00      1 node054
 
=== Query a specific active job: scontrol ===
Show all the details of a currently active job, so not a completed job.
<source lang='bash'>
login ~]$ scontrol show jobid 4241
JobId=4241 Name=WB20F06
  UserId=megen002(16795409) GroupId=domain users(16777729)
  Priority=1 Account=(null) QOS=normal
  JobState=RUNNING Reason=None Dependency=(null)
  Requeue=1 Restarts=0 BatchFlag=1 ExitCode=0:0
  RunTime=02:55:25 TimeLimit=3-08:00:00 TimeMin=N/A
  SubmitTime=2013-12-09T13:37:29 EligibleTime=2013-12-09T13:37:29
  StartTime=2013-12-09T13:37:29 EndTime=2013-12-12T21:37:29
  PreemptTime=None SuspendTime=None SecsPreSuspend=0
  Partition=research AllocNode:Sid=login0:21799
  ReqNodeList=(null) ExcNodeList=(null)
  NodeList=node023
  BatchHost=node023
  NumNodes=1 NumCPUs=4 CPUs/Task=1 ReqS:C:T=*:*:*
  MinCPUsNode=1 MinMemoryNode=0 MinTmpDiskNode=0
  Features=(null) Gres=(null) Reservation=(null)
  Shared=OK Contiguous=0 Licenses=(null) Network=(null)
  Command=/lustre/scratch/WUR/ABGC/...
  WorkDir=/lustre/scratch/WUR/ABGC/...
</source>
 
=== Check on a pending job ===
A submitted job could result in a pending state when there are not enough resources available to this job.
In this example I sumbit a job, check the status and after finding out is it '''pending''' I'll check when is probably will start.
<source lang='bash'>
[@login jobs]$ sbatch hpl_student.job
Submitted batch job 740338
 
[@login jobs]$ squeue -l -j 740338
Fri Feb 21 15:32:31 2014
  JOBID PARTITION    NAME    USER    STATE      TIME TIMELIMIT  NODES NODELIST(REASON)
740338 ABGC_Stud HPLstude bohme999  PENDING      0:00 1-00:00:00      1 (ReqNodeNotAvail)
 
[@login jobs]$ squeue --start -j 740338
  JOBID PARTITION    NAME    USER  ST          START_TIME  NODES NODELIST(REASON)
740338 ABGC_Stud HPLstude bohme999  PD  2014-02-22T15:31:48      1 (ReqNodeNotAvail)
</source>
So it seems this job will problably start the next day, but's thats no guarantee it will start indeed.
 
== Removing jobs from a list: scancel ==
If for some reason you want to delete a job that is either in the queue or already running, you can remove it using the 'scancel' command. The 'scancel' command takes the jobid as a parameter. For the example above, this would be done using the following code:
<source lang='bash'>
scancel 3401
</source>
 
== Allocating resources interactively: sinteractive ==
sinteractive is a tiny wrapper on srun to create interactive jobs quickly and easily. It allows you to get a shell on one of the nodes, with similar limits as you would do for a normal job. To use it, simply run:
<source lang='bash'>
sinteractive -c <num_cpus> --mem <amount_mem> --time <minutes> -p <partition>
</source>
You will then be presented with a new shell prompt on one of the compute nodes (run 'hostname' to see which!). From here, you can test out code in an interactive fashion as needs be.
 
Be advised though - not filling in the above fields will get you a shell with 1 CPU and 100Mb of RAM for 1 hour. This is useful for quick testing, however.
 
=== sinteractive source ===
<source lang='bash'>
#!/bin/bash
srun "$@" -I60 -N 1 -n 1 --pty bash -i
</source>
 
=== interactive Slurm - using salloc ===
If you don't want your shell to be transported but want a new remote shell, do:
<source lang='bash'>
salloc -p ABGC_Low $SHELL
</source>
Now your shell will stay on the login node, but you can do:
<source lang='bash'>
srun <command> &
</source>
To submit tasks to this new shell!
 
Be aware that the time limit of salloc is default 1 hour. If you intend to run jobs for longer times than this, you need to edit the settings for it. See: https://computing.llnl.gov/linux/slurm/salloc.html
 
== Get overview of past and current jobs: sacct ==
To do some accounting on past and present jobs, and to see whether they ran to completion, you can do:
<source lang='bash'>
sacct
</source>
This should provide information similar to the following:
 
        JobID    JobName  Partition    Account  AllocCPUS      State ExitCode
  ------------ ---------- ---------- ---------- ---------- ---------- --------
  3385        BOV-WUR-58  research                    12  COMPLETED      0:0
  3385.batch        batch                                1  COMPLETED      0:0
  3386        BOV-WUR-59  research                    12 CANCELLED+      0:0
  3386.batch        batch                                1  CANCELLED    0:15
  3528        BOV-WUR-59      ABGC                    16    RUNNING      0:0
  3529        BOV-WUR-60      ABGC                    16    RUNNING      0:0
 
Or in more detail for a specific job:
<source lang='bash'>
sacct --format=jobid,jobname,comment,partition,ntasks,alloccpus,elapsed,state,exitcode -j 4220
</source>
This should provide information about job id 4220:
 
      JobID    JobName    Comment  Partition  NTasks  AllocCPUS    Elapsed      State ExitCode
  ------------ ---------- ---------- ---------- -------- ---------- ---------- ---------- --------
  4220        PreProces+              research                  3  00:30:52  COMPLETED      0:0
  4220.batch        batch                              1          1  00:30:52  COMPLETED      0:0
 
'''Job Status Codes'''
 
Typically your job will be either in the Running state of PenDing state. However here is a breakdown of all the states that your job could be in.
 
{| class="wikitable"
|-
!Code!!State!!Description
|-
|CA ||CANCELLED|| Job was explicitly cancelled by the user or system administrator. The job may or may not have been initiated.
|-
|CD|| COMPLETED|| Job has terminated all processes on all nodes.
|-
|CF|| CONFIGURING|| Job has been allocated resources, but are waiting for them to become ready for use (e.g. booting).
|-
|CG|| COMPLETING|| Job is in the process of completing. Some processes on some nodes may still be active.
|-
|F|| FAILED|| Job terminated with non-zero exit code or other failure condition.
|-
|NF|| NODE_FAIL|| Job terminated due to failure of one or more allocated nodes.
|-
|PD|| PENDING|| Job is awaiting resource allocation.
|-
|R|| RUNNING|| Job currently has an allocation.
|-
|S|| SUSPENDED|| Job has an allocation, but execution has been suspended.
|-
|TO|| TIMEOUT|| Job terminated upon reaching its time limit.
|-
|-
|}
 
== Running MPI jobs on Anunna ==
 
[[MPI_on_B4F_cluster | Main article: MPI on Anunna]]
< text here >
 
== Understanding which resources are available to you: sinfo ==
By using the 'sinfo' command you can retrieve information on which 'Partitions' are available to you. A 'Partition' using SLURM is similar to the 'queue' when submitting using the Sun Grid Engine ('qsub'). The different Partitions grant different levels of resource allocation. Not all defined Partitions will be available to any given person. E.g., Master students will only have the Student Partition available, researchers at the ABGC will have 'student', 'research', and 'ABGC' partitions available. The higher the level of  resource allocation, though, the higher the cost per compute-hour. The default Partition is the 'student' partition. A full list of Partitions can be found from the Bright Cluster Manager webpage.
 
<source lang='bash'>
sinfo
</source>
 
  PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
  student*    up  infinite    12  down* node[043-048,055-060]
  student*    up  infinite    50  idle fat[001-002],node[001-042,049-054]
  research    up  infinite    12  down* node[043-048,055-060]
  research    up  infinite    50  idle fat[001-002],node[001-042,049-054]
  ABGC        up  infinite    12  down* node[043-048,055-060]
  ABGC        up  infinite    50  idle fat[001-002],node[001-042,049-054]
 
== See also ==
* [[Tariffs | Costs associated with resource usage]]
* [[Tariffs | Costs associated with resource usage]]
* [[B4F_cluster | Anunna]]
* [[BCM_on_B4F_cluster | BCM on Anunna]]
* [[SLURM_Compare | SLURM compared to other common schedulers]]
* [[Setting_up_Python_virtualenv | Setting up and using a virtual environment for Python3 ]]


== External links ==
== External links ==
* [http://slurm.schedmd.com Slurm official documentation]
* [http://slurm.schedmd.com Slurm official documentation]
* [http://en.wikipedia.org/wiki/Simple_Linux_Utility_for_Resource_Management Slurm on Wikipedia]
* [http://en.wikipedia.org/wiki/Simple_Linux_Utility_for_Resource_Management Slurm on Wikipedia]
* [http://www.youtube.com/watch?v=axWffyrk3aY Slurm Tutorial on Youtube]

Latest revision as of 09:48, 16 June 2026

The resource allocation and scheduling software on Anunna is SLURM: Simple Linux Utility for Resource Management. This page is the entry point — most topics have their own page; below is a short summary plus links.

What's on which page

Quality of Service

When submitting a job, you may optionally assign a different Quality of Service (QoS) to it: 

#SBATCH --qos=std

The QoS values configured on Anunna:

  • std (priority 10) — the default. Use this unless you have a specific reason to pick another.
  • low (priority 1) — reduced priority, but limited to 8 hours per job so a flood of low-priority jobs cannot lock up the cluster.
  • high (priority 20) — higher priority than std. More expensive — see Tariffs.
  • interactive (priority 100) — the highest priority, exclusively for immediate-running interactive jobs. You may not submit many or large jobs at this QoS.

Jobs can in principle be restarted and rescheduled if a higher-priority job needs cluster resources, but at the time of writing this preemption is not actually configured.

Running MPI jobs

For multi-node MPI workloads see MPI on Anunna.

See also

Retrieved from "https://wiki.anunna.wur.nl/index.php?title=Scheduler_Overview_(Slurm)&oldid=2749"