Everything You Need to Know about Using Slurm on Quest

The following is an example batch job submission script. Please note any placeholders which are denoted with <>. You will need to replace these with the appropriate value, and remove the <>. This job submission script would be saved in a file such as jobscript.sh.

#!/bin/bash
#SBATCH --account=<account> ## Required: your allocation/account name, i.e. eXXXX, pXXXX or bXXXX
#SBATCH --partition=short ## Required: (buyin, short, normal, long, gengpu, genhimem, etc)
#SBATCH --time=00:10:00 ## Required: How long will the job need to run (remember different partitions have restrictions on this parameter)
#SBATCH --nodes=1 ## how many computers/nodes do you need (no default)
#SBATCH --ntasks-per-node=1 ## how many cpus or processors do you need on per computer/node (default value 1)
#SBATCH --mem=1G ## how much RAM do you need per computer/node (this affects your FairShare score so be careful to not ask for more than you need))
#SBATCH --job-name=sample_job ## When you run squeue -u 

The first line of the script loads the bash shell. Only the lines that begin with #SBATCH are interpreted by Slurm at the time of job submission. Any words following ## are treated as comments and ignored by Slurm interpreter. Once Slurm places the job on a compute node, the remainder of the script (everything after the last #SBATCH line) is run. Normally in Bash, # is a comment character which means that anything written after a # is ignored by the Bash interpreter/language. When writing a submission script, however, the Slurm interpreter recognizes #SBATCH as a command.

After the Slurm commands, the rest of the script works like a regular Bash script. You can modify environment variables, load modules, change directories, and execute program commands. Lines in the second half of the script that start with # are comments, such as # A regular comment in Bash in the the example above.

Find a downloadable copy of this example script on GitHub.

After you have written and saved your submission script, you can submit your job. At the command line type

sbatch <name_of_script>

where, in the example above <name_of_script> would be jobscript.sh. Upon submission the scheduler will return your job number:

Submitted batch job 549005

If you have a workflow that accepts or needs the jobid as an input for job monitoring or job dependencies, then you may prefer the return value of your job submission be just the job number. To do this, pass the --parsable argument:

sbatch --parsable <name_of_script>
549005

If there is an error in your job submission script, the job will not be accepted by the scheduler and you will receive an error message right away, for example:

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

If your job submission receives an error, you will need to address the issue and resubmit your job. If no error is received, your job has entered the queue and will run.

To specify the account, please include a -A/--account option in your job submission script:

#SBATCH -A <allocation>

Slurm offers two methods, one short (such as -A) and one verbose (--account) to write most of its setting flags. For verbose methods, an = sign is required after the flag (such as --account=).

To submit jobs to Quest, you must be part of an active classroom, research, or buy-in allocation. Additionally, for buy-in allocations, you must be part of a buy-in allocation with access to compute resources. To determine the names of the allocation(s)/account(s) that you are part of on Quest, you can run the following on the command line which will produce output similar to below.

$ groups
quest_demo b1011 p30157 e31572

Once you determine the allocation(s)/account(s) that you are part of, you can check whether the allocation is active and has access to compute resources by running checkproject and reading the last two lines of the output. An example of running checkproject is below.

$ checkproject <allocation> 

==================================== 
Reporting for project pXXXXX
------------------------------------
1 GB in 4623 files (0% of 1000 GB quota)
Allocation Type: Allocation I
Expiration Date: 2022-12-01
Status: ACTIVE
Compute and storage allocation - when status is ACTIVE, this allocation has compute node access and can submit jobs
------------------------------------
====================================

For more information on allocations and allocation management, please see Managing an Allocation on Quest.

To specify the partition, please include a -p/--partition option in your job submission script:

#SBATCH -p <partition>

Quest offers several partitions or queues where you can run your job. Based on the duration of your job, number of cores, and type of access to Quest, you should select the most appropriate partition for your job. A partition must be specified when you submit your job otherwise the scheduler will return the error, "sbatch: error: Batch job submission failed: No partition specified or system default partition".

Partition Definitions: General Access ("p" and "e" accounts)

Standard compute node access.

Specialty compute node access

Partition Definitions: Full Access (buy-ins, or "b" accounts)

Notes

Additional specialized partitions exist for specific allocations. You may be instructed to use a partition name that is not listed above.

If you have a general access allocation and need to run jobs longer than one week, contact Research Computing for a consultation. Some special accommodations can be made for jobs requiring the resources of up to a single node for a month or less.

To specify the walltime for your job, please include a -t/--walltime option in your job submission script:

#SBATCH -t <timelimit>

There are two important considerations when selecting the walltime, the partition that you chose and how long your job is expected to run. Although the partition that you choose will control the maximum wall time that can be selected, we do not recommend to simply select the maximum allowable wall time for that partition unless it is truly needed. There are two problematic ways to set wall time that will create undesired outcomes:

• If your walltime is much longer than your job needs to run, then your job will take longer to start running.
• If your walltime is shorter than your job needs to run, then your job will fail as there is no way to extend the walltime of a running job on Quest.

This is why we recommend submitting a single, representative job and seeing how long it takes to run before selecting a walltime and submitting a large number of jobs. Please note that Slurm considers only the actual duration of your job (not the defined walltime) when calculating your resource utilization.

To specify the number of nodes, please include the -N/--nodes option in your job submission script:

#SBATCH --nodes=<number_of_nodes>

Although the number of nodes is an optional setting, we strongly recommend always setting this value. Specifically, we recommend setting this value to

#SBATCH --nodes=1

as the vast majority of software can only run on a single computer and cannot run across multiple computers. When you forget to set this value, but you do set the Number of Cores, this can cause Slurm to match you with a set of computing resources which your application will be unable to use, but will still be penalized in your fair share for using. If you know that your application uses Message Passing Interface (MPI) to parallelize, then setting this value to something >1 could make sense.

There are two methods for specifying the number of cores, the -n/--ntasks option which indicates how many total cores you would like:

#SBATCH --ntasks=<number_of_cores>

or the --ntasks-per-node= option whichindicates how many cores you would like per node and should always be used with the Number of Nodes option:

#SBATCH --nodes=<number_of_nodes>
#SBATCH --ntasks-per-node=<number_of_cores_per_node>

Although the number of cores is an optional setting whose default is 1, we strongly recommend always setting this value. Specifically, we recommend setting this value (to start) to

#SBATCH --ntasks=1

The only situation in which -n/--ntasks should be greater than 1 is if the application you are using has the capability to be parallelized. Many applications do not have this capability and therefore it is best to start of setting this value to 1. If your application is capable of parallelization, you will next want to determine what type of parallelization it uses in order to set this value correctly. For instance, if your application utilizes shared memory parallelization (OpenMP, R's doParallel, Python's multiprocessing, MATLAB local parpool, etc) then you can consider setting this value to be greater than 1. However, shared memory parallelization can only utilize CPUs within a single computer and CPUs allocated across computers will go unused. Therefore, if your code is parallelized in this manner, you must also specify

#SBATCH --nodes=1

If you know that your application uses Message Passing Interface (MPI) to parallelize, then it can utilize CPUs allocatedacross computers and therefore setting -n/--ntasks without also setting -N/--nodes would make sense.

A final consideration when selecting the number of CPUs is how many CPUs are available on each of the different generations/families of compute nodes that make up Quest. Below is a table which summarizes the relevant information.

To drive home this point, imagine you made the following request:

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=41
#SBATCH --partition=short

This request would eliminate Slurm's ability to match you with any of the computers from generation quest9 and would increase the amount of time it will take to schedule your job as only one type of compute node is able to match your request.

There are two methods for specifying how much memory/RAM you need, the --mem option which indicates how much memory you wantper node.

#SBATCH --mem=<memory per node>G

or the --mem-per-cpu option which indicates how much memory/RAM you need per CPU.

#SBATCH --mem-per-cpu=<memory per cpu>G

If your job submission script does not specify how much memory your job requires, then the default setting is 3.25 GB of memory per core.

#SBATCH --mem-per-cpu=3256M

Therefore, you submitted a job to run on 10 cores and did not specify your memory request in your job submission script, Slurm will allocate 32.5 GB in total.

The memory that is allocated to your job via this setting creates a hard upper limit and your application cannot access memory beyond what Slurm reserves for them; if your job tries to access more memory than has been reserved, it will be terminated.

There is a special setting to request the entire memory of the computer.

#SBATCH --mem=0

How much memory this ends up being will depend on what generation/family of computer Slurm matches you to. The following is a table which summarizes the relevant information.

A final consideration when selecting the amount of memory/RAM is the available memory/RAM on each of the different generations/families of compute nodes that make up Quest. To drive home this point, imagine you made the following request:

#SBATCH --nodes=1
#SBATCH --mem=190G
#SBATCH --partition=short

This request would eliminate Slurm's ability to match you with any of the computers from generation quest9 or quest10 and would increase the amount of time it will take to schedule your job as you will have reduced the pool of available compute nodes.

How can I tell if my job needs more memory to run successfully?

Use the sacct -X command to see information about your recent jobs, for example:

$ sacct -X
           JobID    JobName  Partition    Account  AllocCPUS      State ExitCode 
    ------------ ---------- ---------- ---------- ---------- ---------- -------- 
    1273539      lammps-te+      short     p1234          40  COMPLETED      0:0 
    1273543      vasp-open+      short     p1234          40 OUT_OF_ME+    0:125

The "State" field is the status of your job when it finished. Jobs with a "COMPLETED" state have run without system errors. Jobs with an "OUT_OF_ME+" state have run out of memory and failed. "OUT_OF_ME+" jobs need to request more memory in their job submission scripts to complete successfully.

If the job you're investigating is not recent enough to be listed by sacct -X, add date fields to the command to see jobs between specific start and end dates. For example, to see all jobs between September 15, 2019 and September 16, 2019:
 

$ sacct -X --starttime=091519 --endtime=091619

Specify the date using MMDDYY. More information on sacct is available here.

My job ran out of memory and failed, now what?

First, determine how much memory your job needs following the steps outlined below. Once you know how much memory your job needs, edit your job submission script to reserve that amount of memory + 10% for your job.

To find your job's memory utilization on a compute node:

1. create a test job by editing your job's submission script to reserve all of the memory of the node it runs on
2. run your test job
3. confirm your test job has completed successfully
4. use seff to see how much memory your job actually used.

Create a test job

To profile your job's memory usage, create a test job by modifying your job's submission script to include the lines:

#SBATCH --mem=0
#SBATCH --nodes=1

Setting --mem=0 reserves all of the memory on the node for your job; if you already have a --mem= directive in your job submission script, comment it out. Now your job will not run out of memory unless your job needs more memory than is available on the node.

Setting --nodes=1 reserves a single node for your job. For jobs that run on multiple nodes such as MPI-based programs, request the number of nodes that your job utilizes. Be sure to specify a value for #SBATCH --nodes= or the cores your job submission script reserves could end up on as many nodes as cores requested. Be aware that by setting --mem=0, you will be reserving all the memory on each of those nodes that your cores are reserved on.

2) Run your test job

Submit your test job to the cluster with the sbatch command. For interactive jobs, use srun or salloc.

3) Did your test job complete successfully?

When your job has stopped running, use the sacct -X command to confirm your job finished with state "COMPLETED". If your test job finishes with an "OUT_OF_ME+" state, confirm that you are submitting the modified job submission script that requests all of the memory on the node. If the "OUT_OF_ME+" errors persist, your job may require more memory than is available on the compute node it ran on. In this case, please email quest-help@northwestern.edu for assistance.

4) How much memory did your job actually use?

To see how much memory it used run the command: seff <test_job_id_number>. This returns output similar to:

Job ID: 767731
    Cluster: quest
    User/Group: abc123/abc123
    State: COMPLETED (exit code 0)
    Cores: 1
    CPU Utilized: 00:10:00
    CPU Efficiency: 100.00% of 00:10:00 core-walltime
    Job Wall-clock time: 00:10:00
    Memory Utilized: 60.00 GB
    Memory Efficiency: 50.00% of 120.00 GB
    

Check the job State reported in the 4th line. If it is "COMPLETED (exit code 0)", look at the last two lines. "Memory Utilized" is the amount of memory your job used, in this case 60Gb.

If the job State is FAILED or CANCELLED, the Memory Efficiency percentage reported by seff will be extremely inaccurate. The seff command only works on jobs that have COMPLETED successfully.

How much memory should I reserve in my job script?

It's a good idea to reserve slightly more memory than your job utilized since the same job may require slightly different amounts of memory depending on variations in data it processes in each run of the job. To correctly reserve memory for this job, edit your test job submission script to modify the #SBATCH --mem= directive to reserve 10% more than 60Gb in the job submission script:

#SBATCH --mem=66G

For jobs that use MPI, remove the #SBATCH --mem= directive from your job submission script. Now specify the amount of memory you'd like to reserve per core instead. For example, if your job uses 100Gb of memory total and runs on 10 cores, reserve 10Gb plus a safety factor per cpu:

#SBATCH --mem-per-cpu=11G

If it doesn't matter how many nodes your cores are distributed on, you may remove the #SBATCH --nodes= directive as well.

Be careful not to reserve significant amounts of memory beyond what your job requires as your job's wait time will increase. Reserving excessive memory also wastes shared resources that could be used by other researchers.

To specify a file into which both the standard output and standard error from your job will be written, please include only the -o/--output option in your job submission script:

#SBATCH --output=<name of file>.out

This will cause a file to be created in the submission directory with the name you defined. You may also specify filename which includes the absolute or full path to the file but you cannot just include a path to a directory. Please make sure that all directories in the file path name exist on Quest.

To separate out the standard output and standard error into two separate files, please include both the -o/--output option and the -e/--error option in your job submission script:

#SBATCH --output=<name of file>.out
#SBATCH --error=<name of file>.err

If you do not include either option, the default setting will be to write both the standard output and standard error from your job in a file called

slurm-<slurm jobid>.out

where <slurm jobid> is the ID given to your job by SLURM. You can replicate this default naming scheme yourself by providing the following option:

#SBATCH --output=slurm-%j.out

In addition to %j which will add the job id to the name of your output file, there is also %x which will add the job name to the name of your output file.

To specify a name for your job, please include a -J/--job-name option in your job submission script:

#SBATCH --job-name=<job name>

The default is to set this value to the name of the submission file, so we recommend that you set this to a memorable string because it can be useful when trying to identify a specific job among several running or completed jobs.

To receive e-mails regarding the status of your Slurm jobs, please include both the --mail-type option and the --mail-user option in your job submission script:

#SBATCH --mail-type=<job state that triggers email> ## BEGIN, END, FAIL or ALL
#SBATCH --mail-user=<email address>

If you do not include both of these options, then you will not receive emails from Slurm. Also, you can include any combination of BEGIN, END, FAIL as an argument for this option.

To specify an architecture constraint for your job, please include a -C/--constraint option in your job submission script:

#SBATCH --constraint=<name of compute node architecture>

Not all Quest compute nodes are the same. We currently have four different generations or architectures of compute nodes which we refer to as quest9, quest10, quest11, and quest12 and a summary table of these architectures is provided below.

You can find detailed information on each of these architectures here. If you need to restrict your job to a particular architecture, you can do so through the constraint directive. For example, --constraint=quest10 will cause the scheduler to only match you to compute nodes of the quest10 generation.

Moreover, if you would like to match to any generation of compute nodes, but would like all the compute nodes to be either of generation quest9 or quest10 or quest11 or quest12 and not a combination of generations, then you can set the following for constraint.

#SBATCH --constraint="[quest9|quest10|quest11|quest12]"

This is recommended for jobs that are parallelized using MPI.

The variables in the following table are set by Slurm and they are accessbile in the environment of your job after the job has started running

The squeue command can be used display information about your current jobs on Quest.

The sacct command can be used display information about your past and current jobs on Quest.

Starting with Slurm 22.05, there is a new feature of sacct which allows users to query the job submission script for a given Slurm job for up to 1 year. The syntax is as follows:

sacct -j <job_num> -B

Please note that the -j option is required to query the submission script. Please find an example of this use of sacct below.

$ sacct -j 1454461 -B
Batch Script for 1454461
--------------------------------------------------------------------------------
#!/bin/bash
#SBATCH -A a9009
#SBATCH -p all
#SBATCH --gres=gpu:p100:1
#SBATCH --output=anaconda_pytorch.txt
#SBATCH --job=anaconda_pytorch
#SBATCH -N 1
#SBATCH -n 1
#SBATCH -t 0:10:00
#SBATCH --mem=20G
#SBATCH --nodelist=qgpu6038

module purge all
module load python-miniconda3/4.12.0
source activate /projects/intro/envs/pytorch-1.10-py38
python training_pytorch.py
source deactivate
source activate /projects/intro/envs/pytorch-1.11-py38
python training_pytorch.py

Additionally, sacct can be used to provide specific information about your Slurm jobs. Below, we demonstrate the default/basic behavior of sacct. By default, sacct will only display information about jobs from today and only a subset of information: the jobid, jobname, partition, account, AllocCPUS, state, and exitcode.

$ sacct -X -u <netid>
JobID      JobName Partition  Account AllocCPUS   State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
1453894      bash    all   a9009     1 COMPLETED   0:0
1454434   sample_job    all   a9009     52   FAILED   6:0 

The standard output of sacct may not provide the information we want. To remedy this, we can use the --format flag to choose what we want in our output. The format flag is handled by a list of comma separated variables which specify output data:

$ sacct --user=<netid> --format=var_1,var_2, ... ,var_N

A chart of some variables is provided below:

In addition, only jobs from today will be displayed. To change this, you can supply a start time and an end time via the --starttime and --endtime flags, respectively. For example, suppose you want to find information about jobs that were run on September 12, 2022 and you want to show information regarding the job name, the number of nodes used in the job, the number of cpus, and the elapsed time. Your command would look like this:

$ sacct -X --starttime=09/12/22 --format=jobname,nnodes,ncpus,elapsed

A full list of variables that specify data handled by sacct can be found with the --helpformat flag or by visiting the slurm page on sacct.

To see the maximum amount of memory used/needed by your job run the command: seff <job_id>. This returns output similar to:

Job ID: 767731
    Cluster: quest
    User/Group: abc123/abc123
    State: COMPLETED (exit code 0)
    Cores: 1
    CPU Utilized: 00:10:00
    CPU Efficiency: 100.00% of 00:10:00 core-walltime
    Job Wall-clock time: 00:10:00
    Memory Utilized: 60.00 GB
    Memory Efficiency: 50.00% of 120.00 GB

Check the job State reported in the 4th line. If it is "COMPLETED (exit code 0)", look at the last two lines. "Memory Utilized" is the amount of memory your job used, in this case 60Gb.

If the job State is FAILED or CANCELLED, the Memory Efficiency percentage reported by seff will be extremely inaccurate. The seff command only works on jobs that have COMPLETED successfully.

The checkjob command displays detailed information about a submitted job’s status and diagnostic information that can be useful for troubleshooting submission issues. It can also be used to obtain useful information about completed jobs such as the allocated nodes, resources used, and exit codes.

Example usage:

checkjob <JobID>

where you can get your <JobID> using the squeue commands above.

Example for a Successfully Running Job

[abc123@quser21 ~]$ checkjob 548867
--------------------------------------------------------------------------------------------------------------------
JOB INFORMATION
--------------------------------------------------------------------------------------------------------------------
JobId=548867 JobName=high-throughput-cpu_000094
   UserId=abc123(123123) GroupId=abc123(123) MCS_label=N/A
   Priority=1315 Nice=0 Account=p12345 QOS=normal
   JobState=RUNNING Reason=None Dependency=(null)
   Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0
   RunTime=00:13:13 TimeLimit=00:40:00 TimeMin=N/A
   SubmitTime=2019-01-22T12:51:42 EligibleTime=2019-01-22T12:51:43
   AccrueTime=2019-01-22T12:51:43
   StartTime=2019-01-22T15:52:20 EndTime=2019-01-22T16:32:20 Deadline=N/A
   PreemptTime=None SuspendTime=None SecsPreSuspend=0
   LastSchedEval=2019-01-22T15:52:20
   Partition=short AllocNode:Sid=quser21:15454
   ReqNodeList=(null) ExcNodeList=(null)
   NodeList=qnode[5056-5060]
   BatchHost=qnode5056
   NumNodes=5 NumCPUs=120 NumTasks=120 CPUs/Task=1 ReqB:S:C:T=0:0:*:*
   TRES=cpu=120,mem=360G,node=5,billing=780
   Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*
   MinCPUsNode=1 MinMemoryCPU=3G MinTmpDiskNode=0
   Features=(null) DelayBoot=00:00:00
   OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)
   Command=(null)
   WorkDir=/projects/p12345/high-throughput
   StdErr=/projects/p12345/high-throughput/lammps.error
   StdIn=/dev/null
   StdOut=/projects/p12345/high-throughput/lammps.output
   Power=
--------------------------------------------------------------------------------------------------------------------
JOB SCRIPT
--------------------------------------------------------------------------------------------------------------------
#!/bin/bash
#SBATCH --account=p12345
#SBATCH --partition=normal
#SBATCH --job-name=high-throughput-cpu
#SBATCH --ntasks=120
#SBATCH --mem-per-cpu=3G
#SBATCH --time=00:40:00
#SBATCH --error=lammps.error
#SBATCH --output=lammps.output

module purge
module load lammps/lammps-22Aug18

mpirun -n 120 lmp -in in.fcc

Note in the output above that:

• The JobState is listed as RUNNING.
• The time passed after job start and the total walltime request are given with RunTime and TimeLimit.
• The node name(s) are listed after NodeList.
• The paths to job's working directory (WorkDir), standard error (StdErr) and output (StdOut) files are given.
• If a batch job script is used for submission, the script is presented at the end.

You can cancel one or all of your jobs with scancel. Proceed with caution, as this cannot be undone, and you will not be prompted for confirmation after issuing the command.

Users can place their jobs in a "JobHeldUser" state while submitting the job or after the job has been queued. Running jobs cannot be placed on hold.

The job status will be shown in the output of monitoring commands such as squeue or checkjob.

To release a job from user hold state:

scontrol release <JobID>

The job control command (scontrol) can also be used for changing the parameters of a submitted job before it starts running. The following parameters can be modified safely:

• Job dependency (change to "none")
• Partition (queue)
• Job name
• Wall clock limit
• Allocation

The table below contains some useful examples of using scontrol to change a job's parameters.

For a complete listing of scontrol options, see the official scontrol documentation.

Slurm implements a multi-factor priority scheme for ordering the queue of jobs waiting to be run. sprio command is used to see the contribution of different factors to a pending job's scheduling priority.

For running jobs, you can see the starting priority using checkjob <jobID> command.

Submitting an Interactive Job (to run an application without Graphical User Interface - GUI)

To launch an interactive job from the Quest log-in node in order to run an application without a GUI, use either the srun or salloc command. If you use srun to run an interactive job, then SLURM will automatically launch a terminal session on the compute node after it schedules the job and you simply need to wait for this to happen. Due to the behavior of srun, if you lose connection to your interactive session, the interactive job will terminate.

[quser23 ~]$srun -N 1 -n 1 --account=<account> --mem=<memory>G --partition=<partition> --time=<hh:mm:ss> --pty bash -l
srun: job 3201233 queued and waiting for resources
srun: job 3201233 has been allocated resources
----------------------------------------
srun job start: Mon Mar 14 13:25:41 CDT 2022
Job ID: 3201233
Username: <netid>
Queue: <partition>
Account: <account>
----------------------------------------
The following variables are not
guaranteed to be the same in
prologue and the job run script
----------------------------------------
PATH (in prologue) : /hpc/usertools:/usr/lib64/qt-3.3/bin:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/usr/lpp/mmfs/bin:/opt/ibutils/bin
WORKDIR is: /home/<netid>
----------------------------------------
[qnode0114 ~]$

If you use salloc instead, it will not automatically launch a terminal session on the compute node. Instead, after it schedules your job/request, it will tell you the name of the compute node at which point you can run ssh qnodeXXXX to directly connect to the compute node. Due to the behavior of salloc, if you lose connection to your interactive session, the interactive job will not terminate.

[quser21 ~]$ salloc -N 1 -n 1 --account=<account> --mem=&ltmemory>G --partition=<partition> --time=<hh:mm:ss>
salloc: Pending job allocation 276305
salloc: job 276305 queued and waiting for resources
salloc: job 276305 has been allocated resources
salloc: Granted job allocation 276305
salloc: Waiting for resource configuration
salloc: Nodes qnode8029 are ready for job
[quser21 ~]$ ssh qnode8029
Warning: Permanently added 'qnode8029,172.20.134.29' (ECDSA) to the list of known hosts.
[qnode8029 ~]$

Submitting an Interactive Job (to run an application with Graphical User Interface)

To launch an interactive job from the Quest log-in node in order to run an application with a GUI, first you need to connect to Quest using an application with X11 forwarding support. We recommend using the FastX3 client. Once you have connected to Quest with X11 forwarding enabled, you can then use either the srun or salloc command. If you use srun to run an interactive job, then SLURM will automatically launch a terminal session on the compute node after it schedules the job and you simply need to wait for this to happen. Due to the behavior of srun, if you lose connection to your interactive session, the interactive job will terminate.

[quser23 ~]$srun --x11 -N 1 -n 1 --account=<account> --mem=<memory>G --partition=<partition> --time=<hh:mm:ss> --pty bash -l
srun: job 3201233 queued and waiting for resources
srun: job 3201233 has been allocated resources
----------------------------------------
srun job start: Mon Mar 14 13:25:41 CDT 2022
Job ID: 3201233
Username: <netid>
Queue: <partition>
Account: <account>
----------------------------------------
The following variables are not
guaranteed to be the same in
prologue and the job run script
----------------------------------------
PATH (in prologue) : /hpc/usertools:/usr/lib64/qt-3.3/bin:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/usr/lpp/mmfs/bin:/opt/ibutils/bin
WORKDIR is: /home/<netid>
----------------------------------------
[qnode0114 ~]$

If you use salloc instead, it will not automatically launch a terminal session on the compute node. Instead, after it schedules your job/request, it will tell you the name of the compute node at which point you can run ssh qnodeXXXX to directly connect to the compute node. Due to the behavior of salloc, if you lose connection to your interactive session, the interactive job will not terminate.

[quser21 ~]$ salloc --x11 -N 1 -n 1 --account=<account> --mem=<XXG> --partition=<partition> --time=<hh:mm:ss>
salloc: Pending job allocation 276305
salloc: job 276305 queued and waiting for resources
salloc: job 276305 has been allocated resources
salloc: Granted job allocation 276305
salloc: Waiting for resource configuration
salloc: Nodes qnode8029 are ready for job
[quser21 ~]$ ssh -X qnode8029
Warning: Permanently added 'qnode8029,172.20.134.29' (ECDSA) to the list of known hosts.
[qnode8029 ~]$

Job arrays can be used to submit multiple jobs at once that use the same application script. This can be useful if you want to run the same script multiple times with different input parameters.

In the example below, the --array option defines the job array, with a specification of the index numbers you want to use (in this case, 0 through 9). The $SLURM_ARRAY_TASK_ID bash environmental variable takes on the value of the job array index for each job (so here, integer values 0 through 9, one value for each job). In this example, the value of $SLURM_ARRAY_TASK_ID is used to select the correct index from the input_args bash array which was constructed by reading in input_args.txt, each row of which is then passed on to a script as command line arguments.

jobsubmission.sh

#!/bin/bash
#SBATCH --account=w10001  ## YOUR ACCOUNT pXXXX or bXXXX
#SBATCH --partition=w10001  ### PARTITION (buyin, short, normal, w10001, etc)
#SBATCH --array=0-9 ## number of jobs to run "in parallel" 
#SBATCH --nodes=1 ## how many computers do you need
#SBATCH --ntasks-per-node=1 ## how many cpus or processors do you need on each computer
#SBATCH --time=00:10:00 ## how long does this need to run (remember different partitions have restrictions on this param)
#SBATCH --mem-per-cpu=1G ## how much RAM do you need per CPU (this affects your FairShare score so be careful to not ask for more than you need))
#SBATCH --job-name="sample_job_\${SLURM_ARRAY_TASK_ID}" ## use the task id in the name of the job
#SBATCH --output=sample_job.%A_%a.out ## use the jobid (A) and the specific job index (a) to name your log file
#SBATCH --mail-type=ALL ## you can receive e-mail alerts from SLURM when your job begins and when your job finishes (completed, failed, etc)
#SBATCH --mail-user=email@u.northwestern.edu  ## your email

module purge all
module load python-anaconda3
source activate /projects/intro/envs/slurm-py37-test

IFS=$'\n' read -d '' -r -a input_args < input_args.txt

python slurm_test.py --filename ${input_args[$SLURM_ARRAY_TASK_ID]}

where input_args.txt contains the following:

input_args.txt

filename1.txt
filename2.txt
filename3.txt
filename4.txt
filename5.txt
filename6.txt
filename7.txt
filename8.txt
filename9.txt
filename10.txt

and myscript.py contains the following code:

myscript.py

import argparse
import time

def parse_commandline():
    """Parse the arguments given on the command-line.
    """
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--filename",
                       help="Name of file",
                       default=None)


    args = parser.parse_args()

    return args


###############################################################################
# BEGIN MAIN FUNCTION
###############################################################################
if __name__ == '__main__':
    args = parse_commandline()
    #time.sleep(10) # Sleep for 3 seconds
    print(args.filename)

In this example, myscript.py will receive the values in input.csv as arguments: the first field will be sys.argv[1], the second field will be sys.argv[2], etc.

Note: make sure the number you specify for the --array parameter matches the number of lines in your input file!

Also, note that in this example standard output and error files are printed separately for each element of the job array with the --output and --error options. To avoid each element overwriting these files, tag them with jobID (%A) and elementID (%a) variables (which are automatically assigned by the scheduler) so elements have their own distinct output and error files. An example of this is shown in the jobsubmission.sh presented above.

Submit this script with:

sbatch jobsubmission.sh

The job array will then be submitted to the scheduler with each array element requesting the same resources (such as number of cores, time, memory etc.) as defined in the job submission script.

Dependent jobs are a series of jobs which run or wait to run conditional on the state of another job. For instance, you may submit two jobs and you want the first job to complete successfully before the second job runs. In order to submit this type of workflow, you pass sbatch the jobid of the job that needs to finish before this job starts via the command line argument:

--dependency=afterok:<jobid>

To accomplish this, it is helpful to write all of your sbatch commands in bash script. You will notice that anything you can tell slurm via #SBATCH in the submission script itself, you can also pass to sbatch via the command line. The key here is that the bash variable jid0, jid1, jid2 will contain the jobid that SLURM assigns after you run the sbatch command.

wrapper_script.sh

#!/bin/bash
jid0=($(sbatch --time=00:10:00 --account=w10001 --partition=w10001 --nodes=1 --ntasks-per-node=1 --mem=8G --job-name=example --output=job_%A.out example_submit.sh))

echo "jid0 ${jid0[-1]}" >> slurm_ids

jid1=($(sbatch --dependency=afterok:${jid0[-1]} --time=00:10:00 --account=w10001 --partition=w10001 --nodes=1 --ntasks-per-node=1 --mem=8G --job-name=example --output=job_%A.out --export=DEPENDENTJOB=${jid0[-1]} example_submit.sh))

echo "jid1 ${jid1[-1]}" >> slurm_ids

jid2=($(sbatch --dependency=afterok:${jid1[-1]} --time=00:10:00 --account=w10001 --partition=w10001 --nodes=1 --ntasks-per-node=1 --mem=8G --job-name=example --output=job_%A.out --export=DEPENDENTJOB=${jid1[-1]} example_submit.sh))

echo "jid2 ${jid2[-1]}" >> slurm_ids

In the above, the second job will not start until the first job is finished and the third job will not start until the second one is finished. The actual submission script that is being run is below.

example_submit.sh

#!/bin/bash
#SBATCH --mail-type=ALL ## you can receive e-mail alerts from SLURM when your job begins and when your job finishes (completed, failed, etc)
#SBATCH --mail-user=email@u.northwestern.edu ## your email

if [[ -z "${DEPENDENTJOB}" ]]; then
    echo "First job in workflow"
else
    echo "Job started after " $DEPENDENTJOB
fi

module purge all
module load python-anaconda3
source activate /projects/intro/envs/slurm-py37-test

python --version
python myscript.py --job-id $DEPENDENTJOB

where myscript.py contains the following code:

myscript.py

import argparse
import time


def parse_commandline():
    """Parse the arguments given on the command-line.
    """
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--job-id",
                       help="Job number",
                       default=0)

    args = parser.parse_args()

    return args


###############################################################################
# BEGIN MAIN FUNCTION
###############################################################################
if __name__ == '__main__':
    args = parse_commandline()
    time.sleep(3) # Sleep for 3 seconds
    print(args.job_id)

In this example, we print the job id that had to finish in order for the dependent job to begin. Therefore, the very first job should print 0 because it did not rely on any job to finish in order to run but the second job should print the jobid of the first job and so on.

bash wrapper_script.sh

This will submit the three jobs in sequence and you should see jobs 2 and 3 pending for reason DEPENDENCY.

Total priority score is combination of several factors. These factors are the following:

i. Fair Share Quest's job scheduler uses a fair share mechanism to dynamically determine a score. The calculation is based on the comparison between your share of the resources and your actual usage of these resources. If you or other members of your allocation used large amounts of resources in the recent past, the priority of current jobs will be lower. Accordingly, your jobs will wait longer before they start if the scheduler queue is busy.

On the other hand, if the job queue is empty and the compute resources are idling, regardless of the priority, your jobs will run. You will never run out of compute hours in this model.

Fairshare also includes a recovery mechanism for job priority. The contribution of past resource usage to priority calculations decays over time. Without new usage, the job scores will be restored significantly within a month.

If you are using a general access allocation, the fair share scores of your jobs will be affected from the overall resource usage by all members of the allocation. If you are using a buy-in allocation that has its own compute nodes, your own usage of the those nodes will be the determining factor for your job's fair share score.

ii. Allocation Type There are two types of allocations for research projects on Quest. Research I allocations are ideal for small to moderate computational needs whereas Research II allocations require considerably more resources. Due to this difference in computational needs, a Research II allocation has a higher share of the system resources (or initial fairshare score) compared to a Research I allocation. This is equivalent to receiving more compute hours with a Research II than a Research I.

iii. Job Age Age is length of time an eligible job has been waiting in the queue. Jobs will accumulate priority proportional to their age. This can help overcome starting priority differences between jobs coming from other actors.

iv. Partition A priority score is associated with partition that the job uses. This is only applicable for certain buy-in allocations which have multiple partitions.

More detailed information about Slurm's multi-factor priority system, please see this page.

There is a secondary mechanism that starts lower priority jobs on slots reserved by higher priority jobs while these jobs are acquiring their full set of resources. This is called "backfill" scheduling which helps to increase the utilization of the compute nodes and guarantees no delay in starting the higher priority jobs. To benefit from this mechanism, it is important to accurately request resources (wall time, core, memory) so that the scheduler can find appropriate space on the resource map. Please review resource utilization page for different methods you can use to identify your job's needs.

From time to time, compute resources cannot be backfilled effectively and nodes/cores may appear idle while jobs are waiting on the queue.

If your job submission script generates an error when you submit it with the sbatch command, the problem in your script is in one or more of the lines that begin with #SBATCH. To debug job scripts that generate errors, look up the error message in the section below to identify the most likely reason your script received that error message. Once you have identified the mistake in your script, edit your script to correct it and re-submit your job. If you receive the same error message again, examine the error message and the mistake in your script more closely. Sometimes the same error message can be generated by two different mistakes in the same script, meaning it is possible that you may resolve the first mistake but need to correct a second mistake to clear that particular error message. Mistakes can be difficult to identify, and often require careful reading of your #SBATCH lines.

When you re-submit your job you may receive a new error message. This means the mistake that generated the first error message has been resolved, and now you need to fix another mistake. Slurm returns up to two distinct error messages at a time. If your submission script has more than two mistakes, you will need to re-submit your job multiple times to identify and fix all of them.

When Slurm encounters a mistake in your job submission script, it does not read the rest of your script that comes after the mistake. If the mistake generates an error, you can fix it and resubmit your job, however not all mistakes generate errors. If your script's required elements (account, partition, nodes, cores, and wall time) have been read successfully before Slurm encounters your mistake, your job will still be accepted by the scheduler and run, just not the way you expect it to. Scripts with mistakes that don't generate errors still need to be debugged since the scheduler has ignored some of your #SBATCH lines. You can identify a script with mistakes if the output from your job is unexpected or incorrect.

To use the references below, search for the exact error message generated by your job. Some error messages appear to be similar but are generated by different mistakes.

Note that the errors listed in this document may also be generated by interactive job submissions using srun or salloc. In those cases, the error messages will begin with srun error or salloc error. The procedure to resolve these error messages is the same.

With certain combinations of GUI editors and character sets on your personal computer, copying and pasting into Quest job submission scripts may bring in specific hidden characters that interfere with the scheduler's ability to interpret the script. In these cases, #SBATCH lines will have no mistakes but still generate errors when submitted to the scheduler. To see all of the hidden characters in your job submission script, use the command cat -A <script_name>. To resolve this, you may need to type your submission script into a native unix editor like vi and not use copy and paste.

sinfo -o "%g %.10R %.20l"
GROUPS      PARTITION         TIMELIMIT
b1234       buyin             168:00:00

#SBATCH --account=b1234
#SBATCH --partition=buyin
#SBATCH --time=168:00:00

sinfo -o "%g %.10R %.20l"
GROUPS      PARTITION         TIMELIMIT
b1234       buyin             168:00:00

#SBATCH --account=b1234
#SBATCH --partition=buyin
#SBATCH --time=168:00:00

sinfo -o "%g %.10R %.20l"
GROUPS      PARTITION         TIMELIMIT
b1234       buyin             168:00:00

Error generated sbatch: unrecognized option ‘--n-tasks-per-node=1'

#SBATCH --ntasks-per-node=1

sinfo -o "%g %.10R %.20l %.10c"
GROUPS      PARTITION       TIMELIMIT       CPUS
b1234       buyin           2-00:00:00      20+

#SBATCH --ntasks-per-node=20

Once your job has been accepted, the Slurm scheduler will return a job id number. After waiting in the queue, your job will run. To see the status of your job, use the command sacct -X.

For jobs with mistakes that do not give error messages, you will need to investigate if you notice something is wrong with how the job runs. If you notice a problem on the list below, click on it for debugging suggestions.

module purge all

#SBATCH –-output=/path/to/file/file_name

#SBATCH --error=/path/to/file/file_name

#SBATCH –-output=/path/to/file/"%x.o%j"

#SBATCH –-error=/path/to/file/"%x.e%j"

Besides errors in your script or hardware failure, your job may be aborted by the system if it is still running when the walltime limit you requested (or the upper walltime limit for the partition) is reached. You will see TIMEOUT state for these jobs.

If you use more cores than you requested, the system will again stop the job. This can happen with programs that are multi-threaded. Similarly, if the job exceeds the requested memory, the job will be terminated. Due to this, it is important to profile your code for the memory requirement.

If you do not set the number of nodes/cores, memory or time in your job submission script, the default values will be assigned by the scheduler.

Your job could also fail if you exceed your storage quote in your home or projects directory.

Check how much space you are using in your home directory with

homedu

or

du -h --max-depth=0 ~

Check how much space is used in your projects directory with

checkproject <allocationID>