Skip to content

SLURM

Slurm (Simple Linux Utility for Resource Management) is an open-source workload manager and job scheduler designed for high-performance computing clusters. It is widely used in research, academia, and industry to efficiently manage and allocate computing resources such as CPUs, GPUs, memory, and storage for running various types of jobs and tasks. Slurm helps optimize resource utilization, minimizes job conflicts, and provides a flexible framework for distributing workloads across a cluster of machines. It offers features like job prioritization, fair sharing of resources, job dependencies, and real-time monitoring, making it an essential tool for orchestrating complex computational workflows in diverse fields.

Availability

Software Module Load Command
slurm module load wulver

Please note that the module wulver is already loaded when a user logs in to the cluster. If you use module purge command, make sure to use module load wulver in the slurm script to load SLURM.

Application Information, Documentation

The documentation of SLURM is available at SLURM manual.

Managing and Monitoring Jobs

SLURM has numerous tools for monitoring jobs. Below are a few to get started. More documentation is available on the SLURM website.

The most common commands are:

  • List all current jobs: squeue
  • Job deletion: scancel [job_id]
  • Run a job: sbatch [submit script]
  • Run a command: srun <slurm options> <command name>

SLURM User Commands

Task Command
Job submission: sbatch [script_file]
Job deletion: scancel [job_id]
Job status by job: squeue [job_id]
Job status by user: squeue -u [user_name]
Job hold: scontrol hold [job_id]
Job release: scontrol release [job_id]
List enqueued jobs: squeue
List nodes: sinfo -N OR scontrol show nodes
Cluster status: sinfo

Using SLURM on Wulver

In Wulver, SLURM submission will have new requirements, intended for a more fair sharing of resources without impinging on investor/owner rights to computational resources. All jobs must now be charged to a PI-group (Principal Investigator) account.

  1. To specify the job, use --account=PI_ucid. You can specify --account as either an sbatch or #SBATCH parameter. If you don't know the UCID of PI, usesacctmgr show user $LOGNAME, and you can find the PI's UCID under Def Acct column.

   [ab1234@login01 ~]$ sacctmgr show user $LOGNAME
      User   Def Acct    Admin
---------- ----------  ---------
    ab1234     xy1234      None
2. Wulver has three partitions, differing in GPUs or RAM available:

Partition Nodes Cores/Node CPU GPU Memory Service Unit (SU) Charge
--partition=general 100 128 2.5G GHz AMD EPYC 7763 (2) NA 512 GB 1 SU per hour per cpu
--partition=debug 1 4 2.5G GHz AMD EPYC 7763 (2) NA 512 GB No charges, must be used with --qos=debug
--partition=gpu 25 128 2.0 GHz AMD EPYC 7713 (2) NVIDIA A100 GPUs (4) 512 GB 3 SU per hour per GPU node
--partition=bigmem 2 128 2.5G GHz AMD EPYC 7763 (2) NA 2 TB 1.5 SU per CPU hour
3. Wulver has three levels of “priority”, utilized under SLURM as Quality of Service (QoS):
QoS Purpose Rules Wall time limit, hours Valid Users
--qos=standard Normal jobs, similar to Lochness “public” access SU charges based on node type (see partitions table above), jobs can be preempted by high QoS enqueued jobs 72 Everyone
--qos=low Free access, no SU charge jobs can be preempted by high or standard QoS enqueued jobs 72 Everyone
--qos=high Only available to owners/investors Highest Priority Jobs, no SU Charges. 72 owner/investor PI Groups
--qos=debug Intended for debugging and testing jobs No SU Charges, maximum 4 CPUs are allowed, must be used with --partition=debug 8 Everyone
4. Check Quota Faculty PIs are allocated 300,000 Service Units (SU) per year upon request at no cost, which can be utilized via --qos=standard on the SLURM job. It's important to regularly check the usage of SUs so that users can be aware of their consumption and switch to --qos=low to prevent exhausting all allocated SUs. Users can check their quota using the quota_info UCID command. 
1
2
3
4
5
6
7
8
9
[ab1234@login01 ~]$ module load wulver
[ab1234@login01 ~]$ quota_info $LOGNAME
Usage for account: xy1234
   SLURM Service Units (CPU Hours): 277557 (300000 Quota)
   PROJECT Storage: 867 GB (of 2048 GB quota)
     User ab1234 Usage: 11 GB (No quota)
   SCRATCH Storage: 791 GB (of 10240 GB quota)
     User ab1234 Usage: 50 GB (No quota)
HOME Storage ab1234 Usage: 0 GB (of 50 GB quota)
Here, 'xy1234' represents the UCID of the PI, and 'SLURM Service Units (CPU Hours): 277557 (300000 Quota)' indicates that members of the PI group have already utilized 277,557 CPU hours out of the allocated 300,000 SUs. Please ensure that you load the 'wulver' module before running the 'quota_info' command. This command also displays the storage usage of directories such as $HOME, /project, and /scratch. Users can view both the group usage and individual usage of each storage. In the given example, the group usage from the 2TB project quota is 867 GB, with the user's usage being 11 GB out of that 867 GB. For more details file system quota, see Wulver Filesystem.

Example of slurm script

Submitting Jobs on CPU Nodes

Sample Job Script to use: submit.sh 
    #!/bin/bash -l
    #SBATCH --job-name=job_nme
    #SBATCH --output=%x.%j.out # %x.%j expands to slurm JobName.JobID
    #SBATCH --error=%x.%j.err
    #SBATCH --partition=general
    #SBATCH --qos=standard
    #SBATCH --account=PI_ucid # Replace PI_ucid which the NJIT UCID of PI
    #SBATCH --nodes=1
    #SBATCH --ntasks-per-node=8
    #SBATCH --time=59:00  # D-HH:MM:SS
    #SBATCH --mem-per-cpu=4000M
  • Here, the job requests 1 node with 8 cores, on the general partition with qos=standard. Please note that the memory relies on the number of cores you are requesting.
  • As per the policy, users can request up to 4GB memory per core, therefore the flag --mem-per-cpu is used for memory requirement.
  • In this above script --time indicates the wall time which is used to specify the maximum amount of time that a job is allowed to run. The maximum allowable wall time depends on SLURM QoS, which you can find in QoS.
  • To submit the job, use sbatch submit.sh where the submit.sh is the job script. Once the job has been submitted, the jobs will be in the queue, which will be executed based on priority-based scheduling.
  • To check the status of the job use squeue -u $LOGNAME and you should see the following 
      JOBID PARTITION     NAME     USER  ST    TIME    NODES  NODELIST(REASON)
   635   general     job_nme   ucid   R   00:02:19    1      n0088
    Here, the ST stands for the status of the job. You may see the status of the job ST as PD which means the job is pending and has not been assigned yet. The status change depends upon the number of users using the partition and resources requested in the job. Once the job starts, you will see the output file with an extension of .out. If the job causes any errors, you can check the details of the error in the file with the .err extension.

Submitting Jobs on GPU Nodes

In case of submitting the jobs on GPU, you can use the following SLURM script

Sample Job Script to use: gpu_submit.sh 
    #!/bin/bash -l
    #SBATCH --job-name=gpu_job
    #SBATCH --output=%x.%j.out # %x.%j expands to slurm JobName.JobID
    #SBATCH --error=%x.%j.err
    #SBATCH --partition=gpu
    #SBATCH --qos=standard
    #SBATCH --account=PI_ucid # Replace PI_ucid which the NJIT UCID of PI
    #SBATCH --nodes=1
    #SBATCH --ntasks-per-node=8
    #SBATCH --gres=gpu:2
    #SBATCH --time=59:00  # D-HH:MM:SS
    #SBATCH --mem-per-cpu=4000M

This will request 2 GPUS per node on the GPU partition.

Submitting Jobs on debug

The "debug" QoS in Slurm is intended for debugging and testing jobs. It usually provides a shorter queue wait time and quicker job turnaround. Jobs submitted with the "debug" QoS have access to a limited set of resources (Only 4 CPUS on Wulver), making it suitable for rapid testing and debugging of applications without tying up cluster resources for extended periods.

Sample Job Script to use: debug_submit.sh 
    #!/bin/bash -l
    #SBATCH --job-name=debug
    #SBATCH --output=%x.%j.out # %x.%j expands to slurm JobName.JobID
    #SBATCH --error=%x.%j.err
    #SBATCH --partition=debug
    #SBATCH --qos=debug
    #SBATCH --account=PI_ucid # Replace PI_ucid which the NJIT UCID of PI
    #SBATCH --nodes=1
    #SBATCH --ntasks-per-node=4
    #SBATCH --time=7:59:00  # D-HH:MM:SS, Maximum allowable Wall Time 8 hours
    #SBATCH --mem-per-cpu=4000M

Interactive session on a compute node

Interactive sessions are useful for tasks that require direct interaction with the compute node's resources and software environment. To start an interactive session on the compute node, use the following after logging into Wulver

   srun -p general -n 1 --ntasks-per-node=8 --qos=standard --account=PI_ucid --mem-per-cpu=2G --time=59:00 --pty bash

   srun -p gpu -n 1 --ntasks-per-node=8 --qos=standard --account=PI_ucid --mem-per-cpu=2G --gres=gpu:2 --time=59:00 --pty bash

   srun -p debug -n 1 --ntasks-per-node=4 --qos=debug --account=PI_ucid --mem-per-cpu=2G --gres=gpu:2 --time=59:00 --pty bash

Replace PI_ucid with PI's NJIT UCID.

Warning

Login nodes are not designed for running computationally intensive jobs. You can use the head node to edit and manage your files, or to run small-scale interactive jobs. The CPU usage is limited per user on the head node. Therefore, for serious computing either submit the job using sbatch command or start an interactive session on the compute node.

Note

Please note that if you are using GPUs, check whether your script is parallelized. If your script is not parallelized and only depends on GPU, then you don't need to request more cores per node. In that case use --ntasks-per-node=1, as this will request 1 CPU per GPU. It's important to keep in mind that using multiple cores on GPU nodes may result in unnecessary CPU hour charges. Additionally, implementing this practice can make service unit accounting significantly easier.

Additional Resources