Running Jobs on Swing

Swing, like the other LCRC clusters, utilize the Slurm Workload Manager for job management. Slurm is an open source, fault-tolerant, and highly scalable cluster management and job scheduling system for large and small Linux clusters.

As a cluster workload manager, Slurm has three key functions. First, it allocates exclusive and/or non-exclusive access to resources (compute nodes) to users for some duration of time so they can perform work. Second, it provides a framework for starting, executing, and monitoring work (normally a parallel job) on the set of allocated nodes. Finally, it arbitrates contention for resources by managing a queue of pending work.

The simplest way to become familiar with Slurm and its basic commands is to follow their Quick Start User Guide. In the rest of this page, we’ll cover specific examples and commands. If at any time something becomes unclear, please do contact LCRC support.

Please note that unlike other LCRC clusters, Swing is using Ubuntu as the operating system instead of CentOS.

Allocations Note
Swing, unlike other LCRC clusters, charges allocation time based on GPU Hours instead of Core Hours. Please factor this in when applying for time on Swing.

Please see GPU Hour Usage below for more details.

Logging Into Swing

Please be sure to following our Getting Started documentation in order to make sure you’ve completed the necessary steps so that you can login to Swing. Once you’ve done this, you can SSH to Swing by running the following:

ssh <your_argonne_username>@swing.lcrc.anl.gov

The LCRC login nodes should not be used to run jobs on. Doing so may impact other users and require these login nodes to be rebooted.

If you need to add a new SSH key as you may not have logged in for awhile, please read through our documentation here.

Swing also shares the same global GPFS filesystem as other LCRC clusters. All of your home and project directories noted in our storage documentation will be available between clusters.

Projects Used for Job Submission

LCRC resources require a valid project with an allocation to submit jobs. Projects are what keeps track of your quarterly allocations. Please see the following page for more information about Projects in LCRC.

To see how much time will be deducted from your project when running jobs on Swing, please see the following on GPU Hour Usage.

When logging into Swing for the first time, you’ll want to change your default project (as a reference, what LCRC calls projects are referred to as accounts in Slurm).

All LCRC clusters currently use separate allocation/time databases. Your time and balances on one cluster will not be the same on the other. If you need to check your current account’s (project’s) balance(s), change your default account, etc., please see our documentation below or reference the information here: Project Allocation Queries and Management.

  • All Argonne Employee Swing users will have a default project set to startup-<username> upon first login with a project balance of 100 GPU hours.
  • All Non-Argonne Employee Swing users will have a default project set to external upon first login which has no time allocated and thus you will not be able to submit jobs.

Query your Default Project on Swing

You can check your default project on Swing and make sure this is set correctly with this command:

lcrc-sbank -q default

Setting a Default Project on Swing

You can set your default project on Swing with the following command:

lcrc-sbank -s default <project_name>

You can also specify the project name on Swing in your job submission if you’d like to use something different other than your default. With SBATCH, this can be done with:

#SBATCH -A <project_name>

Query Project Balances on Swing

You can query your project balances on Swing to see how much time you have available and how much you have used.

Query all of your project balances on Swing:

lcrc-sbank -q balance

Query a specific project balance on Swing:

lcrc-sbank -q balance <project_name>

Query a Project Transaction History on Swing

If you’d like to see the transaction history for a project on Swing, you can run the below.

lcrc-sbank -q trans <project_name>

lcrc-sbank Help Menu

If you need to query the lcrc-sbank help menu at any time, simply run the below.

lcrc-sbank -h

Software Environment Using Lmod

Swing is using Lmod (Lua Environment Modules) for environment variable management. Lmod has several advantages. For example, it prevents you from loading multiple versions of the same package at the same time. It also prevents you from having multiple compilers and MPI libraries loaded at the same time. See the Lmod User Guide for information on how to use Lmod.

The software environment is unique to Swing. We have already built some basic software. We will continue to install some standard software, but complex applications should be built by the user.

Please note that Python modules are available in the ‘anaconda3’ module and are not listed separately via Lmod.

Using Slurm to Submit Jobs

Swing is using Slurm for the job resource manager and scheduler for the cluster.

The Slurm Workload Manager is an open source, fault-tolerant, and highly scalable cluster management and job scheduling system for large and small Linux clusters.

As a cluster workload manager, Slurm has three key functions. First, it allocates exclusive and/or non-exclusive access to resources (compute nodes) to users for some duration of time so they can perform work. Second, it provides a framework for starting, executing, and monitoring work (normally a parallel job) on the set of allocated nodes. Finally, it arbitrates contention for resources by managing a queue of pending work.

Your best source of finding information on using Slurm will come from their quickstart guide here or by using the man pages.

Below we will outline some general information on the Swing Slurm partitions and supply some basic submission information to get you started.

Partitions Limits

Swing currently enforces the following limits on publicly available partitions:

  • 32 Running Jobs per user.
  • 100 Queued Jobs per user.
  • 7 Days (168 Hours) Maximum Walltime.
  • 1 Hour Default Walltime if not specified.
  • gpu is the default (and only) partition.

Available Partitions

Swing has only one partition and it is the default, which is named gpu. You do not need to specify this in your submit script. By default, you will be allocated 1/8th of the node resources per GPU (32 cores + 128GB RAM per GPU).

Nodes allow for multiple jobs from multiple users up until the resources are fully consumed (8 jobs with 1 GPU each per node, 1 job with 8 GPUs per node, and everything in between). Slurm will clean up temporary files when all of your jobs on a node exit.

You MUST request at least 1 GPU to run a job otherwise you will see the following error:

srun: error: Please request at least 1 GPU in the partition 'gpu'
srun: error: e.g '#SBATCH --gres=gpu:1')
srun: error: Unable to allocate resources: Invalid generic resource (gres) specification

Here’s a short list of node specifications:


Partition Name Number of Nodes GPUs Per Node GPU Memory Per Node CPUs Per Node DDR4 Memory Per Node Operating System
gpu 6 8x NVIDIA A100 40GB 320GB 2x AMD EPYC 7742 64-Core Processor (128 Total Cores) 1TB Ubuntu 20.04.2 LTS

Job Submission Commands

The 3 most common tools you will use to submit jobs are sbatch, srun and salloc.

You can reference the table below for a simple, quick cheat sheet on a few examples about jobs in Slurm:

Slurm Command Description
sbatch <job_script> Submit <job_script> to the Scheduler
srun <options> Run Parallel Jobs
salloc <options> Request an Interactive Job
squeue View Job Information
scancel <job_id> Delete a Job

Example sbatch Job Submission

Let’s go through a simple compile and job submission using sbatch.

Here, we have a simple Hello World application for GPUs:

[user@gpulogin1 ~]$ cat hello.cu
#include <stdio.h>

#define NUM_BLOCKS 2
#define BLOCK_WIDTH 16

__global__ void hello()
{
  printf("Hello world! I'm thread %d in block %d\n", threadIdx.x, blockIdx.x);
}

int main(int argc, char **argv)
{

  // launch the kernel
  hello<<<NUM_BLOCKS, BLOCK_WIDTH>>>();

  // force the printf()s to flush
  cudaDeviceSynchronize();

  printf("That's all!\n");

  return 0;

}

Load an appropriate module and build your application:

[user@gpulogin1 ~]$ module load nvhpc
[user@gpulogin1 ~]$ nvcc hello.cu -o hello-gpu-run

Let’s take a look at a simple submit script:

[user@gpulogin1 ~]$ cat hello.sh
#!/bin/bash
#
#SBATCH --job-name=gpu-test
#SBATCH --account=<my_lcrc_project_name>
#SBATCH --nodes=2
#SBATCH --gres=gpu:8
#SBATCH --time=00:05:00

srun ./hello-gpu-run

Now, submit your job:

[user@gpulogin1 ~]$ sbatch hello.sh
Submitted batch job 398

You can use squeue to check the status of your job. Once finished, check that your application ran successfully:

[user@gpulogin1 ~]$ cat slurm-398.out
Hello world! I'm thread 0 in block 0
Hello world! I'm thread 1 in block 0
Hello world! I'm thread 2 in block 0
Hello world! I'm thread 3 in block 0
Hello world! I'm thread 4 in block 0
Hello world! I'm thread 5 in block 0
Hello world! I'm thread 6 in block 0
Hello world! I'm thread 7 in block 0
Hello world! I'm thread 8 in block 0
Hello world! I'm thread 9 in block 0
Hello world! I'm thread 10 in block 0
Hello world! I'm thread 11 in block 0
Hello world! I'm thread 12 in block 0
Hello world! I'm thread 13 in block 0
Hello world! I'm thread 14 in block 0
Hello world! I'm thread 15 in block 0
Hello world! I'm thread 0 in block 1
Hello world! I'm thread 1 in block 1
Hello world! I'm thread 2 in block 1
Hello world! I'm thread 3 in block 1
Hello world! I'm thread 4 in block 1
Hello world! I'm thread 5 in block 1
Hello world! I'm thread 6 in block 1
Hello world! I'm thread 7 in block 1
Hello world! I'm thread 8 in block 1
Hello world! I'm thread 9 in block 1
Hello world! I'm thread 10 in block 1
Hello world! I'm thread 11 in block 1
Hello world! I'm thread 12 in block 1
Hello world! I'm thread 13 in block 1
Hello world! I'm thread 14 in block 1
Hello world! I'm thread 15 in block 1
That's all!
Hello world! I'm thread 0 in block 0
Hello world! I'm thread 1 in block 0
Hello world! I'm thread 2 in block 0
Hello world! I'm thread 3 in block 0
Hello world! I'm thread 4 in block 0
Hello world! I'm thread 5 in block 0
Hello world! I'm thread 6 in block 0
Hello world! I'm thread 7 in block 0
Hello world! I'm thread 8 in block 0
Hello world! I'm thread 9 in block 0
Hello world! I'm thread 10 in block 0
Hello world! I'm thread 11 in block 0
Hello world! I'm thread 12 in block 0
Hello world! I'm thread 13 in block 0
Hello world! I'm thread 14 in block 0
Hello world! I'm thread 15 in block 0
Hello world! I'm thread 0 in block 1
Hello world! I'm thread 1 in block 1
Hello world! I'm thread 2 in block 1
Hello world! I'm thread 3 in block 1
Hello world! I'm thread 4 in block 1
Hello world! I'm thread 5 in block 1
Hello world! I'm thread 6 in block 1
Hello world! I'm thread 7 in block 1
Hello world! I'm thread 8 in block 1
Hello world! I'm thread 9 in block 1
Hello world! I'm thread 10 in block 1
Hello world! I'm thread 11 in block 1
Hello world! I'm thread 12 in block 1
Hello world! I'm thread 13 in block 1
Hello world! I'm thread 14 in block 1
Hello world! I'm thread 15 in block 1
That's all!

Assuming everything went well, you’ve successfully compiled and submitted your test job!

Example Interactive Job Submission

There are a couple of ways to run an interactive job:

First, you can just get a session on a node by using the srun command in the following way:

srun --gres=gpu:1 --pty bash

This will drop you onto one node with 1 GPU allocated to you. Once you exit the node, the allocation will be relinquished.


If you want more flexibility, you can instead have the system first allocate resources for the job using the
the salloc command:

salloc -N 2 --gres=gpu:4 -t 00:30:00

This job will allocate 2 nodes with 4 GPUs each for 30 minutes. You should get the job number from the output. This command will not log you into any of your allocated nodes by default.

You can get a list of your allocated nodes and many other slurm settings set by the salloc command by doing:

printenv | grep SLURM

After the resources were allocated and the session was granted use srun command to run your job.

When you allocate resources via salloc, you can also now freely SSH to the nodes in your allocation as well if you prefer to run jobs from the nodes themselves.

Checking Queues and Jobs

To view job and job step information use squeue.

Here’s a quick example of what the output may look like:

squeue
             JOBID PARTITION       NAME    USER ST       TIME  NODES NODELIST(REASON)
               999     gpu  test-joba   user2  R    2:40:31      2 gpu[3-4]
               998     gpu  test-job2   user1  R      45:20      1 gpu1
               997     gpu  test-job1   user1  R       3:04      1 gpu1

Here are also some common options for squeue:

-a Display information about all jobs in all partitions. This is the default when running squeue with no options.
-u <user_list> Request jobs or job steps from a comma separated list of users. The list can consist of user names or user id numbers.
-j <job_id_list> Requests a comma separated list of job IDs to display. Defaults to all jobs.
-l Report more of the available information for the selected jobs or job steps.

Deleting a Job

To delete a job use scancel. This command will take the job id as its argument. Your job id will be given to when you submit the job. You can also retrieve this from the squeue command detailed above.

scancel <job_id>

Other Useful Slurm Commands

scontrol – can be used to report more detailed information about nodes, partitions, jobs, job steps, and configuration.

Common examples:

scontrol show node node-name Shows detailed information about the nodes.
scontrol show partition partition-name Shows detailed information about a specific partition.
scontrol show job job-id Shows detailed information about a specific job or all jobs if no job id is given.
scontrol update job job-id Change attributes of submitted job.

For an extensive list of formatting options please consult scontrol man page.


sinfo – view information about jobs, nodes and partitions located in the Slurm scheduling queue

Common options:

-a, --all Display information about all partitions.
-t, --states <states> Display nodes in a specific state. Example: idle
-i <seconds>, --iterate=<seconds> Print the state on a periodic basis. Sleep for the indicated number of seconds between reports.
-l, --long Print more detailed information.
-n <nodes>, --nodes=<nodes> Print information only about the specified node(s). Multiple nodes may be comma separated or expressed using a node range expression. For example “gpu[1-2]”
-o <output_format>, --format=<output_format> Specify the information to be displayed using an sinfo format string.

For an extensive list of formatting options please consult sinfo man page.


sacct – command displays accounting data for all jobs and job steps and can be used to display the information about the complete jobs.

Common options:

-S, --starttime Select jobs in any state after the specified time.
-E end_time, --endtime=end_time Select jobs in any state before the specified time.

Valid time formats are:

HH:MM[:SS] [AM|PM]
MMDD[YY] or MM/DD[/YY] or MM.DD[.YY]
MM/DD[/YY]-HH:MM[:SS]
YYYY-MM-DD[THH:MM[:SS]]

Example:

# sacct -S2014-07-03-11:40 -E2014-07-03-12:00 -X -ojobid,start,end,state
                  JobID                 Start                  End        State
              --------- --------------------- -------------------- ------------
              2         2014-07-03T11:33:16   2014-07-03T11:59:01   COMPLETED
              3         2014-07-03T11:35:21   Unknown               RUNNING
              4         2014-07-03T11:35:21   2014-07-03T11:45:21   COMPLETED
              5         2014-07-03T11:41:01   Unknown               RUNNING

For an extensive list of formatting options please consult sacct man page.


sprio – view the factors that comprise a job’s scheduling priority.

sprio is used to view the components of a job’s scheduling priority when the multi-factor priority plugin is installed. sprio is a read-only utility that extracts information from the multi-factor priority plugin. By default, sprio returns information for all pending jobs. Options exist to display specific jobs by job ID and user name.

For an extensive list of formatting options please consult sprio man page.

GPU Hour Usage

As mentioned, submitting jobs to Swing requires time allocated to a Project (or what Slurm calls an Account). Our documentation has an extensive write up on this on the following page: Projects in LCRC

Whenever a computing job runs on any computing node, the time the job uses will be counted and recorded as computing used by the associated project. A job must have a project in order to run on the computing nodes and will be assigned to your default project if none has been specified in your job script. ALL jobs submitted via sbatch, srun or salloc will deduct computing GPU hours from your project. Please note that unlike other LCRC clusters, Swing charges time based on GPU hours instead of CPU core hours. You should factor this in when applying for time on Swing.

On Swing, the compute nodes charge as follows for each job:

GPU Nodes
# of Nodes * # GPUs Per Node * Time Used

Compute Node Scratch Space

Swing currently writes all temporary files on the compute nodes to a 20 GB tmpfs at /scratch. Please note that all data will be deleted from this directory once your job completes. You can also change your environments TMPDIR variable in your job script if you want to set an alternate path.

Why Isn’t My Job Running Yet?

If today is NOT LCRC Maintenance Day and you find that your job is in the pending (PD) state after running squeue, Slurm will provide a reason for this shown in the squeue command. Here are a few of the most common reasons your job may not be running.

First, check to the see reason code by querying your job number in Slurm:

squeue -j <job_id>

Then, you can determine why the job has not started by deciphering this sample reason list:

Reason Code Description
AccountNotAllowed The job isn’t using an account that is allowed on the partition.
AssocGrpBillingMinutes The job doesn’t have enough time in the banking account to begin.
BadConstraints The job’s constraints can not be satisfied.
BeginTime The job’s earliest start time has not yet been reached.
Cleaning The job is being requeued and still cleaning up from its previous execution.
Dependency This job is waiting for a dependent job to complete.
JobHeldAdmin The job is held by a system administrator.
JobHeldUser The job is held by the user.
NodeDown A node required by the job is down.
PartitionNodeLimit The number of nodes required by this job is outside of it’s partitions current limits. Can also indicate that required nodes are DOWN or DRAINED.
PartitionTimeLimit The job’s time limit exceeds it’s partition’s current time limit.
Priority One or more higher priority jobs exist for this partition or advanced reservation.
QOSMaxJobsPerUserLimit The job’s QOS has reached its maximum job count for the user at one time.
ReqNodeNotAvail During LCRC Maintenance Day, you may see this reason, otherwise, some node specifically required by the job is not currently available. The node may currently be in use, reserved for another job, in an advanced reservation, DOWN, DRAINED, or not responding. Nodes which are DOWN, DRAINED, or not responding will be identified as part of the job’s “reason” field as “UnavailableNodes”. Such nodes will typically require the intervention of a system administrator to make available.
Reservation The job is waiting its advanced reservation to become available.
Resources The job is waiting for resources to become available.
TimeLimit The job exhausted its time limit.

While this is not every reason code, these are the most common. You can view the full list of Slurm reason codes here.

Application Specific Examples

Command Line Quick Reference Guide

Command Description
sbatch <script_name> Submit a job.
scancel <job_id> Delete a job.
squeue
squeue -u <username>
Show queued jobs via the scheduler.
Show queued jobs from a specific user.
scontrol show job <job_id> Provide a detailed status report for a specified job via the scheduler.
sinfo -t idle Get a list of all free/idle nodes.
lcrc-sbank -q balance <project_name>
lcrc-sbank -q balance
lcrc-sbank -q default
lcrc-sbank -s default <project_name>
lcrc-sbank -q trans <project_name>
Query a specific project balance.
Query all of your project balances.
Query your default project.
Change your default project.
Query all transactions on a project.
lcrc-quota Query your global filesystem disk usage.

Contact Information

Please contact support@lcrc.anl.gov with any questions you may have regarding Swing usage.