Partitions

SLURM scheduler organizes similar sets of nodes and job features into individual groups, which are individually referred to as a “partition”. Each partition has hard limits for the maximum wall clock time, maximum job size, the upper limit to the number of nodes a user can request, etc. The partitions available for all ISAAC-NG users include “campus”, “campus-gpu”, “campus-bigmem”, “campus-gpu-bigmem”, “long”, “long-gpu”, “long-bigmem”, and “short” under which a number of CPU and GPU cores are available. The default partition is “campus”; in order to use any other partition, your SLURM job must explicitly request it. For more information on available resources under each partition, please refer to our System Overview page.

Scheduling Policy

ISAAC-NG cluster uses a variety of mechanisms to determine how to schedule jobs. Most of the resource request mechanisms can be manipulated by users to ensure that their jobs are scheduled and executed in a reasonable time period.

sacctmgr show qos \
format=name,maxwall,maxtres,maxtrespu,maxjobspu,maxjobspa,maxsubmitpa

 #SBATCH --account ACF-UTK0011 (or #SBATCH -A ACF-UTK0011)
 #SBATCH --partition=campus (or #SBATCH -p campus)
 #SBATCH --qos=campus (or #SBATCH -q condo)

#SBATCH --account ACF-UTK<project-number>
#SBATCH --partition=campus-gpu
#SBATCH --qos=utk
#SBATCH --gpus=<n>

#SBATCH --account ACF-UTK<project-number>
#SBATCH --partition=condo-<condo-name>
#SBATCH --qos=condo

 $ sinfo -Nel

$ scontrol show PartitionName=<partition-name> | grep AllowAccounts

$ sacctmgr -p show assoc | grep <netid>

Slurm Commands

In the table below, are listed few important Slurm commands (with a description of the command) which are most often used while working with Slurm scheduler, and can be used on the ISAAC-NG login nodes.

Slurm Variables

Below are tabulated few important Slurm variables which ISAAC-NG users may find useful.

SBATCH Flags

The jobs a user submits to the ISAAC-NG cluster are submitted using a sbatch command which passes the request for the resources a user has requested in the job script to the Slurm scheduler. The resources in the job script are requested using the “SBATCH” directive. Note that SLURM scheduler can accept SBATCH directives in two formats. Users can choose to use either of the two formats at their own discretion. The description of each of the SBATCH flags is given below:

Submitting Jobs with Slurm

On ISAAC-NG, batch jobs can be submitted in two ways: (i) interactive batch mode, and (ii) non-interactive batch mode.

Interactive Batch mode

Interactive batch jobs give users the interactive access to compute nodes. In interactive batch mode, users can request that the Slurm scheduler allocate the resources of compute nodes, directly on the terminal (e.g., using a terminal’s command line interface (CLI)). A common use for interactive batch jobs is to debug a calculation or a program before submitting non-interactive batch jobs for production runs. This section demonstrates how to run interactive jobs through the batch system and provides common usage tips.

Interactive batch mode can be invoked on an ISAAC-NG login node by using the salloc command followed by specific SBATCH flags to request specific resources to which the user would like access. The different SBATCH flags are given in table 3.3.

$ salloc -A projectaccount --nodes=1 --ntasks=1 --partition=campus --time=01:00:00
or 
$ salloc -A projectaccount -N 1 -n 1 -p campus -t 01:00:00

The salloc command interprets the user’s request to the SLURM scheduler, and as a result, allows the user to request the resources. In the example command above, we requested that the Slurm scheduler allocate one node and one CPU core for a total time of 1 hour using the “campus” partition. Note that if the salloc command is executed without specifying the resources (i.e., number of nodes, number of CPUs, tasks, and wall clock time etc.), then the Slurm scheduler will allocate the default resources.

Important Note: The default resources on ISAAC-NG are: one processor (1 CPU on 1 node) located under the campus partition, with a maximum wall clock time of 1 hour.

When the Slurm scheduler allocates the resources, the user who submitted the job gets a message on the terminal (as shown below) containing the information about the jobid and the hostname of the compute node where the resources are allocated. Notice that in the below example the jobid is “1234” and the hostname of the compute node is “nodename.”

 $ salloc --nodes=1 --ntasks=1 --time=01:00:00 --partition=campus
  salloc: Granted job allocation 1234
  salloc: Waiting for resource configuration
  salloc: Nodes nodename are ready for job
 $

Once the the interactive job starts, the user need to ssh to the allocated node or one of the allocated nodes (for the jobs requesting more than one node) and should change their working directories to a Lustre file system project space or scratch space to run the user’s computationally intense applications. It is best to run large user jobs in Lustre scratch space rather than Lustre project space, because there is no size limit to the amount of data that a user may place on Lustre scratch space; however, there is a limit to how much data that a user may place on Lustre project space. Please visit File System for more information. If users have questions about how to best run their interactive jobs, users are welcome to send their questions to HPSC staf by submitting HPSC Service Request.

To run the parallel executable, we recommend using srun followed by the executable as shown below:

 $ srun executable
or
 $ srun -n <nprocs> executable

Important Note: Users do not necessarily need to include the number of processors they will require, before running the parallel executable while calling srun. The Slurm wrapper srun will execute your calculations in parallel on the number of processors requested in the user’s job script. Serial applications (that is, non-parallel applications) can be run both with and without srun.

Non-interactive batch mode

In non-interactive batch mode, the set of resources as well as the commands for the application to be run are written in a text file, which is referred to as a “batch file” or “batch script”. A user’s “batch script” for their particular job is submitted to the Slurm scheduler by using the sbatch command. Batch scripts are very useful for users who want to run productions jobs. This is because batch scripts allow users to work on a cluster non-interactively. In batch jobs, users submit a group of commands to Slurm and then simply check the job’s status and the output of the commands from time to time. However, sometimes it is very useful to run a job interactively (primarily for debugging). Click here to check how to run the batch jobs interactively. A typical example of a non-interactive batch script/job script is given below:

 #!/bin/bash
 #This file is a submission script to request the ISAAC resources from Slurm 
 #SBATCH -J job			       #The name of the job
 #SBATCH -A ACF-UTK0011              # The project account to be charged
 #SBATCH --nodes=1                     # Number of nodes
 #SBATCH --ntasks-per-node=48          # cpus per node 
 #SBATCH --partition=campus            # If not specified then default is "campus"
 #SBATCH --time=0-01:00:00             # Wall time (days-hh:mm:ss)
 #SBATCH --error=job.e%J	       # The file where run time errors will be dumped
 #SBATCH --output=job.o%J	       # The file where the output of the terminal will be dumped
 #SBATCH --qos=campus

 # Now list your executable command/commands.
 # Example for code compiled with a software module:
 module load example/test

 hostname
 sleep 100
 srun executable

Stuff

The above job script can be divided into three sections:

1. Shell interpreter (one line)
   • The first line of the script specifies the script’s interpreter. The syntax of this line is #!/bin/shellname (sh, bash, csh, ksh, zsh)
   • This line is important and essential. If not mentioned, then scheduler will print an error.
2. SLURM submission options
   • The second section contains a bunch of lines starting with ‘#SBATCH’.
   • These lines are not the comments, but by format start with ‘#’.
   • #SBATCH is a Slurm directive which communicates information regarding the resources requested by the user in the batch script file.
   • #SBATCH options after the first non-comment line are ignored by Slurm scheduler
   • The description about each of the flags is mentioned in the table 3.3
   • The command sbatch on the terminal is used to submit the non-interactive batch script.
3. Shell commands
   • The shell command follows the last #SBATCH line.
   • These commands are a set of commands or tasks which a user wants to run. This also includes any software modules which may be needed to access a particular application.
   • To run the parallel application, it is recommended to use srun followed by the name of the full path and name of the executable if the executable path is not loaded into Slum environment while submitting the script.

For the quick start, we have also provided a collection of complete sample job scripts that are available on the ISAAC-NG cluster at /lustre/isaac/examples/jobs

Checking for Available Resources

In order to best utilize resources and avoid jobs waiting in queue, users should check what Slurm resources are available before deciding job submission parameters and submitting jobs. The Slurm commands sinfo and isaac-sinfo are offered for users to query the system for the status of available resources.

The command sinfo outputs the status of available nodes organized by partition and node-type while isaac-sinfo is a modified version to include partitions directly usable by ISAAC users with cleaner output. If a user was searching for a GPU node to run a GPU-accelerated job, they may for example execute the following:

$ isaac-sinfo | grep gpu | egrep 'idle|mixed'

The above outputs nodes in GPU partitions that are either idle and totally free or are mixed where only some resources are taken by other users’ jobs. This filters out nodes in the ‘allocated’ state since those nodes’ resources are entirely utilized by current running jobs. A sample output line may look like the following:

campus-gpu               intel,cascadelake,avx512    up 30-00:00:0 1-infinite   no       NO        all      1       mixed             clrv0701

Utilizing the information given, users can tailor their Slurm job requests to the available resources. For example, knowing that there resources left available on node “clrv0701” based on the previous output of isaac-sinfo showing ‘mixed’, a user may request the following:

$ salloc --account=ACF-UTK0011 --nodes=1 --ntasks=8 --partition=campus-gpu --qos=campus-gpu --time=1:00:00 --gpus=1 --nodelist=clrv0701

It is important to notice that the amount of nodes, cores, and GPUs requested match what is currently available, based on the output of the aforementioned commands, otherwise the job will wait in queue until the requested resources become available. To ensure a submitted job has requested an appropriate amount of resources based on the desired or available target node(s), users should compare the resources requested with the resources available. Checking available resources applies to all jobs whether they are submitted interactively, non-interactively, or through Open OnDemand.

Job Arrays

Slurm offers a useful option of submitting jobs that use array flags, for users whose batch jobs require identical resources. Using the array flag in a job script, users can submit multiple jobs with with a single sbatch command. Although the job script is submitted only once using sbatch command, the individual jobs in the array (of jobs) are scheduled independently with unique job array identifiers ($SLURM_ARRAY_JOB_ID). Each of the individual jobs can be differentiated using Slurm’s environmental variable $SLURM_ARRAY_TASK_ID. To understand this variable, let us consider an example of a Slurm script given below:

#!/bin/bash
#SBATCH -J myjob
#SBATCH -A ACF-UTK0011
#SBATCH -N 1
#SBATCH --ntasks-per-node=30  ###-ntasks is used when we want to define total number of processors
#SBATCH --time=01:00:00
#SBATCH --partition=campus     #####
##SBATCH -e myjob.e%j   ## Errors will be written in this file
#SBATCH -o myjob%A_%a.out    ## Separate output file will be created for each array. %A will be replaced by jobid and %a will be replaced by array index
#SBATCH --qos=campus
#SBATCH --array=1-30
       # Submit array of jobs numbered 1 to 30
###########   Perform some simple commands   ########################
set -x
###########   Below code is used to create 30 script files needed to submit the array of jobs   ###############
for i in {1..30}; do cp sleep_test.sh 'sleep_test'$i'.sh';done

###########   Run your executable   ###############
sh sleep_test$SLURM_ARRAY_TASK_ID.sh

In the above example, 30 sleep_test$index.sh executable files were created, and these jobs have names are differentiated by an index. One understands that the 30 executable files can be run by either submitting 30 individual jobs or by using Slurm arrays. A Slurm array would take these files in the form of an array named sleep_test[1-30].sh and execute these files. The variable SLURM_ARRAY_TASK_ID is set to array index value [between 1 and 30], with 30 being the total number of jobs in the array in this example, and the index value is defined in the Slurm script (above) using the #SBATCH directive

#SBATCH --array=1-30

The simultaneous number of jobs using a job array can also be limited by using a %n flag along with –array flag. For example: To run only 5 jobs at a time in a Slurm array, users can include the SLURM directive

#SBATCH --array=1-30%5

In order to create a separate output file for each of the submit jobs using Slurm arrays, use %A and %a, which represents the jobid and job array index as shown in the above example.

Exclusive Access to Nodes

As explained in the Scheduling Policy, jobs that are submitted by the same user can share computational nodes. However, if desired, users can request in their job scripts that whole node(s) be reserved to run their jobs without sharing them with other jobs. To do that use the below command:

Interactive batch mode:

 $ salloc -A projectaccount --nodes=1 --ntasks=1 --partition=campus --time=01:00:00 --exclusive --qos=campus

Non-Interactive batch mode:

Add the below line in your job script

 #SBATCH --exclusive

Monitoring Job Status

Users can regularly check status of their jobs by using the squeue command.

$ squeue -u <netID>
              JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
              1202    campus     Job3 username PD       0:00      2 (Resources)
              1201    campus     Job1 username  R       0:05      2 node[001-002]
              1200    campus     Job2 username  R       0:10      2 node[004-005]

The description of each of the columns of the output from squeue command is given below:

When a user submits a job, it passes through various states. The values of these states for a job is given by squeue command under the column ST. The possible values of the job under ST column are given below:

Altering Batch Jobs

The users are allowed to change the attributes of their jobs until the job starts running. In this section, we will describe how to alter your batch jobs with examples.

Remove a Job from Queue

User can remove the jobs in any state which are submitted by them using the command scancel.

To remove a job with a JOB ID 1234, use the command:

scancel 1234

Modifying the Job Details

Users can make use of the Slurm command scontrol which is used to alter a variety of Slurm parameters. Most of the commands using scontrol can only be executed by an ISAAC-NG System Adminstrator; however, users are granted some permissions to use scontrol for use on the jobs they have submitted, provided the jobs are not in the running mode.

Release/Hold a job

scontrol release/hold jobid

Modify the name of the job

scontrol update JobID=jobid JobName=any_new_name

Modify the total number of tasks

scontrol update JobID=jobid NumTasks=Total_tasks

Modify the number of CPUs per node

scontrol update JobID=jobid MinCPUsNode=CPUs

Modify the wall time of the job

scontrol update JobID=jobid TimeLimit=day-hh:mm:ss