FASTER:Batch
FASTER Batch Processing: Slurm
Contents
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 FASTER, 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 FASTER. 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 FASTER (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 FASTER.
| 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.
| 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 FASTER.
| 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 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.
| 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 | --ntasks-per-node=# | --ntasks-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 FASTER 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.
| 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 | $TMPDIR | This is a directory assigned locally on the compute node for the job located at /tmp/job.$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 FASTER: 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 $SCRATCH # Change current directory to /scratch/user/[netID]/ ml purge # Purge all modules ml intel/2020b # Load the intel/2020b module ml # List all currently loaded modules ./myProgram.o # Run "myProgram.o"
For information on the module system or specific software, visit our Modules page and our Software page.
Job Submission
Once you have your job file ready, it is time to submit your job. You can submit your job to slurm with the following command:
[NetID@FASTER1 ~]$ 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 FASTER.
| 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: FASTER 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 FASTER 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 FASTER. 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=Example_SNSC_CPU #Set the job name to "JobExample1" #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --ntasks=1 #Request 1 task #SBATCH --mem=2560M #Request 2560MB (2.5GB) per node #SBATCH --output=Example_SNSC_CPU.%j #Redirect stdout/err to file #SBATCH --partition=cpu #Specify partition to submit job to ##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=Example_SNMC_CPU #Set the job name to Example_SNMC_CPU #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --nodes=1 #Request 1 node #SBATCH --ntasks-per-node=64 #Request 64 tasks/cores per node #SBATCH --mem=248M #Request 248G (248GB) per node #SBATCH --output=Example_SNMC_CPU.%j #Redirect stdout/err to file #SBATCH --partition=cpu #Specify partition to submit job to ##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=Example_MNMC_CPU #Set the job name to Example_MNMC_CPU #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --nodes=2 #Request 2 nodes #SBATCH --ntasks-per-node=64 #Request 64 tasks/cores per node #SBATCH --mem=248G #Request 248G (248GB) per node #SBATCH --output=Example_MNMC_CPU.%j #Redirect stdout/err to file #SBATCH --partition=cpu #Specify partition to submit job to ##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 (single node, single core)
#!/bin/bash ##NECESSARY JOB SPECIFICATIONS #SBATCH --job-name=Example_SNSC_GPU #Set the job name to Example_SNSC_GPU #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --ntasks=1 #Request 1 task #SBATCH --mem=248G #Request 248G (248GB) per node #SBATCH --output=Example_SNSC_GPU.%j #Redirect stdout/err to file #SBATCH --partition=gpu #Specify partition to submit job to #SBATCH --gres=gpu:a100:1 #Specify GPU(s) per node, 1 A100 GPU ##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 (single node, multiple core)
#!/bin/bash ##NECESSARY JOB SPECIFICATIONS #SBATCH --job-name=Example_SNMC_GPU #Set the job name to Example_SNMC_GPU #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --nodes=1 #Request 1 nodes #SBATCH --ntasks-per-node=32 #Request 32 tasks/cores per node #SBATCH --mem=248G #Request 248G (248GB) per node #SBATCH --output=Example_SNMC_GPU.%j #Redirect stdout/err to file #SBATCH --partition=gpu #Specify partition to submit job to #SBATCH --gres=gpu:a100:10 #Specify GPU(s) per node, 10 a100 GPU ##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 (multiple node, multiple core)
#!/bin/bash ##NECESSARY JOB SPECIFICATIONS #SBATCH --job-name=Example_MNMC_GPU #Set the job name to Example_MNMC_GPU #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr 30min #SBATCH --nodes=2 #Request 2 nodes #SBATCH --ntasks-per-node=32 #Request 32 tasks/cores per node #SBATCH --mem=248G #Request 248G (248GB) per node #SBATCH --output=Example_MNMC_GPU.%j #Redirect stdout/err to file #SBATCH --partition=gpu #Specify partition to submit job to #SBATCH --gres=gpu:a100:1 #Specify GPU(s) per node, 1 A100 gpu ##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:
- The total number of jobs that can be concurrently running (number of run slots)
- The wall-clock time limit per job
- 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 August 17, 2022).
| Queue Name | Max Nodes per Job (assoc's cores)* | Max GPUs | Max Duration | Max Jobs in Queue* | Charge Rate (per node-hour) |
|---|---|---|---|---|---|
| development | 1 nodes (64 cores)* | 10 | 1 hr | 1* | 64 Service Unit (SU) + GPUs used |
| cpu | 128 nodes (8,192 cores)* | 0 | 7 days | 50* | 64 Service Unit (SU) |
| gpu | 128 nodes (8,192 cores)* | 10 | 7 days | 50* | 64 Service Unit (SU) + GPUs used |
Checking queue usage
The following command can be used to get information on queues and their nodes.
[NetID@FASTER1 ~]$ sinfo
Example output:
PARTITION AVAIL TIMELIMIT JOB_SIZE NODES(A/I/O/T) CPUS(A/I/O/T) cpu* up 7-00:00:00 1-32 16/61/13/90 1024/3904/832/5760 gpu up 7-00:00:00 1-32 25/5/8/38 844/1076/512/2432 memverge up 2-00:00:00 1-2 0/0/2/2 0/0/128/128 fpga up 2-00:00:00 1-2 0/2/0/2 0/128/0/128
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.
[username@FASTER ~]$ pestat
Example output:
Hostname Partition Node Num_CPU CPUload Memsize Freemem Joblist
State Use/Tot (MB) (MB) JobId User ...
fc004 cpu* idle 0 64 0.01 256000 253490
....
To generate a list of the nodes in the gpu queue and their current configuration:
[username@FASTER ~]$ pestat -p gpu -G
Example output:
Hostname Partition Node Num_CPU CPUload Memsize Freemem GRES/node Joblist
State Use/Tot (15min) (MB) (MB) JobID User GRES/job ...
fc098 gpu idle 0 64 0.00 256000 252732 gpu:a40:4
fc099 gpu idle 0 64 0.00 256000 253123 gpu:a10:2
fc100 gpu idle 0 64 0.00 256000 252685 gpu:t4:8
....
Checking bad nodes
The following command can be used to view a current list of bad nodes on the machine:
[NetID@FASTER1 ~]$ bad_nodes.sh
The following output is just an example output and users should run bad_nodes.sh not see a current list.
Example output:
% bad_nodes.sh REASON USER TIMESTAMP STATE NODELIST The system board OCP1 PG voltage is outside of range. root 2022-07-11T14:38:07 drained fc152 testing DIMM B5 somebody 2022-09-01T12:31:22 drained* fc017 Not responding slurm 2022-09-19T12:10:47 down* fc001
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 4 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.
