Hprc banner tamu.png

Grace:Batch

From TAMU HPRC
Jump to: navigation, search

Grace Batch Processing: Slurm

Introduction

The batch system is a load distribution implementation that ensures convenient and fair use of a shared resource. Submitting jobs to a batch system allows a user to reserve specific resources with minimal interference to other users. All users are required to submit resource-intensive processing to the compute nodes through the batch system - attempting to circumvent the batch system is not allowed.

On Grace, Slurm is the batch system that provides job management. Jobs written in other batch system formats must be translated to Slurm in order to be used on Grace. The Batch Translation Guide offers some assistance for translating between batch systems that TAMU HPRC has previously used.

Building Job Files

While not the only method of submitted programs to be executed, job files fulfill the needs of most users.

The general idea behind job files follows:

  • Make resource requests
  • Add your commands and/or scripting
  • Submit the job to the batch system

In a job file, resource specification options are preceded by a script directive. For each batch system, this directive is different. On Grace (Slurm) this directive is #SBATCH.
For every line of resource specifications, this directive must be the first text of the line, and all specifications must come before any executable lines. An example of a resource specification is given below:

#SBATCH --jobname=MyExample  #Set the job name to "MyExample"

Note: Comments in a job file also begin with a # but Slurm recognizes #SBATCH as a directive.

A list of the most commonly used and important options for these job files are given in the following section of this wiki. Full job file examples are given below.

Basic Job Specifications

Several of the most important options are described below. These basic options are typically all that is needed to run a job on Grace.

Basic Grace (Slurm) Job Specifications
Specification Option Example Example-Purpose
Wall Clock Limit --time=[hh:mm:ss] --time=05:00:00 Set wall clock limit to 5 hours 00 min
Job Name --job-name=[SomeText] --job-name=mpiJob Set the job name to "mpiJob"
Total Task/Core Count --ntasks=[#] --ntasks=96 Request 96 tasks/cores total
Tasks per Node I --ntasks-per-node=# --ntasks-per-node=48 Request exactly (or max) of 48 tasks per node
Memory Per Node --mem=value[K|M|G|T] --mem=360G Request 360 GB per node
Combined stdout/stderr --output=[OutputName].%j --output=mpiOut.%j Collect stdout/err in mpiOut.[JobID]

It should be noted that Slurm divides processing resources as such: Nodes -> Cores/CPUs -> Tasks

A user may change the number of tasks per core. For the purposes of this guide, each core will be associated with exactly a single task.

Additional Grace (Slurm) Job Specifications Warning: these options are NOT COMPATIBLE with OpenMPI
Reset Env I --export=NONE Do not propagate environment to job
Reset Env II --get-user-env=L Replicate the login environment

Example batch file template for an MPI job. Notice that the Reset Env options have been omitted, so this example can work with OpenMPI.


#!/bin/bash 
#                                                    
##NECESSARY JOB SPECIFICATIONS                                  
#SBATCH --job-name=mpiJob                                      
#SBATCH --time=5:00                                            
#SBATCH --ntasks=96                                             
#SBATCH --ntasks-per-node=48                                    
#SBATCH --mem=360G                                             
#SBATCH --output=mpiOut.%j                                     

## YOUR COMMANDS BELOW                                          


Optional Job Specifications

A variety of optional specifications are available to customize your job. The table below lists the specifications which are most useful for users of Grace.

Optional Grace/Slurm Job Specifications
Specification Option Example Example-Purpose
Set Allocation --account=###### --account=274839 Set allocation to charge to 274839
Email Notification I --mail-type=[type] --mail-type=ALL Send email on all events
Email Notification II --mail-user=[address] --mail-user=howdy@tamu.edu Send emails to howdy@tamu.edu
Specify Queue --partition=[queue] --partition=gpu Request only nodes in gpu subset
Specify General Resource --gres=[resource]:[count] --gres=gpu:1 Request one GPU per node
Specify A100 GPU Resource --gres=gpu:[a100]:[count] --gres=gpu:a100:1 Request one a100 GPU per node
Specify RTX 6000 GPU Resource --gres=gpu:[rtx]:[count] --gres=gpu:rtx:2 Request two RTX 6000 GPUs per node
Specify T4 GPU Resource --gres=gpu:t4:[count] --gres=gpu:t4:4 Request four T4 GPUs per node
Submit Test Job --test-only Submit test job for Slurm validation
Request Temp Disk --tmp=M --tmp=10240 Request at least 10 GB in temp disk space
Request License --licenses=[LicenseLoc] --licenses=nastran@slurmdb:12

Alternative Specifications

The job options within the above sections specify resources with the following method:

  • Cores and CPUs are equivalent
  • 1 Task per 1 CPU desired
  • You specify: desired number of tasks (equals number of CPUs)
  • You specify: desired number of tasks per node (equal or less than the 28 cores per compute node)
  • You get: total nodes equal to #ofCPUs/#ofTasksPerNodes
  • You specify: desired Memory per node

Slurm allows users to specify resources in units of Tasks, CPUs, Sockets, and Nodes.

There are many overlapping settings and some settings may (quietly) overwrite the defaults of other settings. A good understanding of Slurm options is needed to correctly utilize these methods.

Alternative Memory/Core/Node Specifications
Specification Option Example Example-Purpose
Node Count --nodes=[min[-max]] --nodes=4 Spread all tasks/cores across 4 nodes
CPUs per Task --cpus-per-task=# --cpus-per-task=4 Require 4 CPUs per task (default: 1)
Memory per CPU --mem-per-cpu=MB --mem-per-cpu=2000 Request 2000 MB per CPU
NOTE: If this parameter is less than 1024, SLURM will misinterpret it as 0
Tasks per Core --ntasks-per-core=# --ntasks-per-core=4 Request max of 4 tasks per core
Tasks per Node II --tasks-per-node=# --tasks-per-node=5 Equivalent to Tasks per Node I
Tasks per Socket --ntasks-per-socket=# --ntasks-per-socket=6 Request max of 6 tasks per socket
Sockets per Node --sockets-per-node=# --sockets-per-node=2 Restrict to nodes with at least 2 sockets

If you want to make resource requests in an alternative format, you are free to do so. Our ability to support alternative resource request formats may be limited.

Using Other Job Options

Slurm has facilities to make advanced resources requests and change settings that most Grace users do not need. These options are beyond the scope of this guide.

If you wish to explore the advanced job options, see the Advanced Documentation.

Environment Variables

All the nodes enlisted for the execution of a job carry most of the environment variables the login process created: HOME, SCRATCH, PWD, PATH, USER, etc. In addition, Slurm defines new ones in the environment of an executing job. Below is a list of most commonly used environment variables.

Basic Slurm Environment Variables
Variable Usage Description
Job ID $SLURM_JOBID Batch job ID assigned by Slurm.
Job Name $SLURM_JOB_NAME The name of the Job.
Queue $SLURM_JOB_PARTITION The name of the queue the job is dispatched from.
Submit Directory $SLURM_SUBMIT_DIR The directory the job was submitted from.
Temporary Directory $TMPDIR This is a directory assigned locally on the compute node for the job located at /work/job.$SLURM_JOBID. Use of $TMPDIR is recommended for jobs that use many small temporary files.

Note: To see all relevant Slurm environment variables for a job, add the following line to the executable section of a job file and submit that job. All the variables will be printed in the output file.

env | grep SLURM

Clarification on Memory, Core, and Node Specifications

Memory Specifications are IMPORTANT.
For examples on calculating memory, core, and/or node specifications on Grace: Specification Clarification.

Executable Commands

After the resource specification section of a job file comes the executable section. This executable section contains all the necessary UNIX, Linux, and program commands that will be run in the job.
Some commands that may go in this section include, but are not limited to:

  • Changing directories
  • Loading, unloading, and listing modules
  • Launching software

An example of a possible executable section is below:

cd $SCRATCH      # Change current directory to /scratch/user/[netID]/ 
ml purge         # Purge all modules
ml intel/2020b   # Load the intel/2020b module
ml               # List all currently loaded modules

./myProgram.o    # Run "myProgram.o"

For information on the module system or specific software, visit our Modules page and our Software page.

Job Submission

Once you have your job file ready, it is time to submit your job. You can submit your job to slurm with the following command:

[NetID@grace1 ~]$ sbatch MyJob.slurm 
Submitted batch job 3606

Job Monitoring and Control Commands

After a job has been submitted, you may want to check on its progress or cancel it. Below is a list of the most used job monitoring and control commands for jobs on Grace.

Job Monitoring and Control Commands
Function Command Example
Submit a job sbatch [script_file] sbatch FileName.job
Cancel/Kill a job scancel [job_id] scancel 101204
Check status of a single job squeue --job [job_id] squeue --job 101204
Check status of all
jobs for a user
squeue -u [user_name] squeue -u User1
Check CPU and memory efficiency for a job
(Use only on finished jobs)
seff [job_id] seff 101204

Here is an example of the seff command provides for a finished job:

% seff 12345678
Job ID: 12345678
Cluster: grace
User/Group: username/groupname
State: COMPLETED (exit code 0)
Nodes: 16
Cores per node: 28
CPU Utilized: 1-17:05:54
CPU Efficiency: 94.63% of 1-19:25:52 core-walltime
Job Wall-clock time: 00:05:49
Memory Utilized: 310.96 GB (estimated maximum)
Memory Efficiency: 34.70% of 896.00 GB (56.00 GB/node)

Job File Examples

Several examples of Slurm job files for Grace are listed below. For translating Ada (LSF) job files, the Batch Job Translation Guide provides some reference.

NOTE: Job examples are NOT lists of commands, but are a template of the contents of a job file. These examples should be pasted into a text editor and submitted as a job to be tested, not entered as commands line by line.

There are several optional parameters available for jobs on Grace. In the examples below, they are commented out/ignored via ##. If you wish to include these values as parameters for your jobs, please change it to a singular # and adjust the parameter value accordingly.

Example Job 1: A serial job (single core, single node)

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample1       #Set the job name to "JobExample1"
#SBATCH --time=01:30:00              #Set the wall clock limit to 1hr and 30min
#SBATCH --ntasks=1                   #Request 1 task
#SBATCH --mem=2560M                  #Request 2560MB (2.5GB) per node
#SBATCH --output=Example1Out.%j      #Send stdout/err to "Example1Out.[jobID]"

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address

#First Executable Line

Example Job 2: A multi core, single node job

#!/bin/bash
##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample2       #Set the job name to "JobExample2"
#SBATCH --time=6:30:00               #Set the wall clock limit to 6hr and 30min
#SBATCH --nodes=1                    #Request 1 node
#SBATCH --ntasks-per-node=8          #Request 8 tasks/cores per node
#SBATCH --mem=8G                     #Request 8GB per node 
#SBATCH --output=Example2Out.%j      #Send stdout/err to "Example2Out.[jobID]" 

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address 

#First Executable Line

Example Job 3: A multi core, multi node job

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample3       #Set the job name to "JobExample3"
#SBATCH --time=1-12:00:00            #Set the wall clock limit to 1 Day and 12hr
#SBATCH --ntasks=8                   #Request 8 tasks
#SBATCH --ntasks-per-node=2          #Request 2 tasks/cores per node
#SBATCH --mem=4096M                  #Request 4096MB (4GB) per node 
#SBATCH --output=Example3Out.%j      #Send stdout/err to "Example3Out.[jobID]"

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address 

#First Executable Line

Example Job 4: A serial GPU job

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample4       #Set the job name to "JobExample4"
#SBATCH --time=01:30:00              #Set the wall clock limit to 1hr and 30min
#SBATCH --ntasks=1                   #Request 1 task
#SBATCH --mem=2560M                  #Request 2560MB (2.5GB) per node
#SBATCH --output=Example4Out.%j      #Send stdout/err to "Example4Out.[jobID]"
#SBATCH --gres=gpu:1                 #Request 1 GPU per node can be 1 or 2
#SBATCH --partition=gpu              #Request the GPU partition/queue

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address 

#First Executable Line

Example Job 5: A serial GPU job with a specific GPU type

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample5       #Set the job name to "JobExample4"
#SBATCH --time=01:30:00              #Set the wall clock limit to 1hr and 30min
#SBATCH --ntasks=1                   #Request 1 task
#SBATCH --mem=2560M                  #Request 2560MB (2.5GB) per node
#SBATCH --output=Example4Out.%j      #Send stdout/err to "Example4Out.[jobID]"
#SBATCH --gres=gpu:rtx:1             #Request 1 "rtx" GPU per node 
#SBATCH --partition=gpu              #Request the GPU partition/queue

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address 

#First Executable Line

Example Job 6: A parallel GPU job

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample6       #Set the job name to "JobExample5"
#SBATCH --time=01:30:00              #Set the wall clock limit to 1hr and 30min
#SBATCH --ntasks=28                   #Request 1 task
#SBATCH --mem=2560M                  #Request 2560MB (2.5GB) per node
#SBATCH --output=Example5Out.%j      #Send stdout/err to "Example5Out.[jobID]"
#SBATCH --gres=gpu:2                 #Request 2 GPU per node can be 1 or 2
#SBATCH --partition=gpu              #Request the GPU partition/queue

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address 

#First Executable Line


See more specialized job files (if available) at the HPRC Software page

Batch Queues

Upon job submission, Slurm sends your jobs to appropriate batch queues. These are (software) service stations configured to control the scheduling and dispatch of jobs that have arrived in them. Batch queues are characterized by all sorts of parameters. Some of the most important are:

  1. The total number of jobs that can be concurrently running (number of run slots)
  2. The wall-clock time limit per job
  3. The type and number of nodes it can dispatch jobs to

These settings control whether a job will remain idle in the queue or be dispatched quickly for execution.

The current queue structure is: (updated on January 11, 2021).

Queue Job Max Cores / Nodes Job Max Walltime Compute Node Types Per-User Limits Across Queues Notes
short 1536 cores / 32 nodes 2 hr 384 GB nodes (800) 6144 Cores per User
medium 6144 cores / 128 nodes 1 day
long 3072 cores / 64 nodes 7 days
xlong 1536 cores / 32 nodes 21 days For jobs needing to run longer than 7 days.

Submit jobs to this partition with the --partition xlong option.

gpu 1536 cores / 32 nodes 4 days A100 GPU nodes (100) Preferred GPU node type if only --gres=gpu:N is used.
Also can request with --gres=gpu:a100:N (N is either 1 or 2)
RTX 6000 GPU nodes (9) Can request with --gres=gpu:rtx:N (N is either 1 or 2)
T4 GPU nodes (8) Can request with --gres=gpu:t4:N (N is 1, 2, 3, or 4)
bigmem 192 cores / 4 node 2 days Large Memory 3TB nodes (8)


Checking queue usage

The following command can be used to get information on queues and their nodes.

[NetID@grace1 ~]$ sinfo

Example output:

PARTITION        AVAIL  TIMELIMIT    JOB_SIZE    NODES(A/I/O/T)   CPUS(A/I/O/T)  
    
short*           up     2:00:00      1-32        32/763/5/800     1496/36664/240/38400    


Note: A/I/O/T stands for Active, Idle, Offline, and Total

Checking node usage

The following command can be used to generate a list of nodes and their corresponding information, including their CPU usage.

[NetID@grace1 ~]$ pestat

Example output:

Hostname       Partition     Node      Num_CPU    CPUload    Memsize    Freemem    Joblist
                             State     Use/Tot               (MB)       (MB)       JobId User ...
c001          short*         idle      0   48     0.01       368640     365067  


Checkpointing

Checkpointing is the practice of creating a save state of a job so that, if interrupted, it can begin again without starting completely over. This technique is especially important for long jobs on the batch systems, because each batch queue has a maximum walltime limit.


A checkpointed job file is particularly useful for the gpu queue, which is limited to 2 days walltime due to its demand. There are many cases of jobs that require the use of gpus and must run longer than two days, such as training a machine learning algorithm.


Users can change their code to implement save states so that their code may restart automatically when cut off by the wall time limit. There are many different ways to checkpoint a job file depending on the software used, but it is almost always done at the application level. It is up to the user how frequently save states are made depending on what kind of fault tolerance is needed for the job, but in the case of the batch system, the exact time of the 'fault' is known. It's just the walltime limit of the queue. In this case, only one checkpoint need be created, right before the limit is reached. Many different resources are available for checkpointing techniques. Some examples for common software are listed below.

Advanced Documentation

This guide only covers the most commonly used options and useful commands.

For more information, check the man pages for individual commands or the Slurm Manual.