GETTING STARTED ON LONGLEAF

Table of Contents

Introduction

System Information

Getting an Account

Logging In

Main Directory Spaces

Applications Environment

Job Submission

Longleaf Partitions

More SLURM Commands

Additional Help

Introduction

Longleaf is a high throughput computing cluster available to researchers, students, faculty and staff across the University. With over 300 nodes, dedicated GPU partitions and highly performant storage, it is optimized for memory- and I/O-intensive, loosely coupled workloads with an emphasis on aggregate job throughput over individual job performance. Longleaf is particularly suited for serial workloads, consisting of many jobs each requiring a single compute host. This contains a longer description of longleaf.

Longleaf users can also access the cluster though Open OnDemand, a web-based portal offering job submission, file system browsing and many preconfigured applications.

High-Level System Information

  • Operating System:
    • Red Hat Enterprise Linux Server release 8
  • Resource management:
    • Job submissions are handled by the Slurm batch processing scheduler

General Partition Nodes

The several general partitions (general/interact, general_big/hov, bigmem, snp, etc.) are comprised of the following nodes:

Node Count CPU Count Memory (GB) GPU Gres
13 384 1500
4 256 991
5 256 927
3 256 734
3 256 732
6 256 500
8 128 991 gpu:nvidia_a100-pcie-40gb:3
1 128 754
5 128 2970
27 72 732
1 64 991 gpu:nvidia_l40s:8
7 56 488 gpu:nvidia_l40:4
40 56 230
99 48 355
19 48 230
3 40 487 gpu:tesla_v100-sxm2-16gb:8
16 40 235 gpu:tesla_v100-sxm2-16gb:4
24 24 235
7 24 112
3 8 41 gpu:nvidia_geforce_gtx_1080:8

Open OnDemand Portal

Open Ondemand is a web-portal that provides a terminal, file browser, and graphical interface for certain apps on the Longleaf cluster.See here for further information.


GPU Hardware and Partitions

Please refer to our page detailing use of GPUs and best practices.



Getting an Account

Follow the steps listed on Request a Cluster Account page and select Longleaf Cluster under subscription type. You will receive an email notification once your account has been created.

Logging In

Linux:

Linux users can use ssh from within their Terminal application to connect to Longleaf.

If you wish to enable x11 forwarding use the “–X” ssh option. Be sure to use your UNC ONYEN and password for the login:

ssh -X <onyen>@longleaf.unc.edu

Windows:

Windows users should download MobaXterm (Home Edition). Then use the Session icon to create a Longleaf SSH session using longleaf.unc.edu for “Remote host” and your ONYEN for the “username” (Port should be left at 22).

Mac:

Mac users can use ssh from within their Terminal application to connect to Longleaf. Be sure to use your UNC ONYEN and password for the login:

ssh -X <onyen>@longleaf.unc.edu

To enable x11 forwarding Mac users will need to download, install, and run Xquartz on their local machine in addition to using the “–X” ssh option. Furthermore, in many instances for x11 forwarding to work properly Mac users need to use the Terminal application that comes with Xquartz instead of the default Mac terminal application.

A successful login takes you to “login node” resources that have been set aside for user access. The login node is where you will edit your code, execute basic UNIX commands, and submit your jobs from to the SLURM job scheduler.

DO NOT RUN YOUR CODE OR RESEARCH APPLICATIONS DIRECTLY ON THE LOGIN NODE. THESE MUST BE SUBMITTED TO SLURM!

In order to connect to Longleaf from an off-campus location, a connection to the campus network through a VPN client is required.

Main Directory Spaces

NAS home space

Your home directory will be in /nas/longleaf/home/<onyen> and is backed up via snapshots.

Your home directory has a quota which you will want to monitor occasionally: 50 GB soft limit and a 75 GB hard limit.

/users storage

Your primary storage directory is: /users/<o>/<n>/<onyen>

This storage is provided by the same hardware as /proj.

High capacity storage

  • OK to compute against, however as IO increases, consider copying or moving to /work for processing
  • OK to hold inactive data sets like a near-line archive
  • If a meaningful amount of cold data accrues, it can be packaged and MOVED to cloud archive, providing more working space for your warm data
  • /users is not intended to be used for team oriented shared storage, like /proj; it is intended to be your personal storage location. Think of it as a capacity expansion to your home directory. In this context, please note that work is NOT intended to be a personal storage location; /work is for data actively being processed with high IO requirements. Please move any data from /work to /users that is not actively being computed upon
  • 10 TB quota

/work storage

Your /work directory will be: /work/users/<o>/<n>/<onyen> (the “o/n/” are the first two letters of your ONYEN).

/work is built for high-throughput and data-intensive computing, and intended for data actively being computed on, accessed and processed by systems such as Longleaf

There is a 10 TB quota per user. For inactive data, please move it to /users or /proj - contact us if you have no place or insufficient quota to move inactive data off of /work.

File systems such as /work in an academic research setting typically employ a file deletion policy, auto-deleting files of a certain age. At this time, there are no time limits for files on /work. We rely upon the community to maintain standards of appropriate and reasonable use of the various storage tiers.

Proj Space

“/proj” space is available to PIs (only) upon request. The amount of /proj space initially given to a PI varies according to their needs. Unlike the former /pine scratch space there is no file deletion policy for /proj space, but users should take care in managing the use of their /proj space to stay under assigned quotas. Note that by default /proj space is not backed up.

For further information and to make a request for /proj space please email research@unc.edu.

Applications Environment

The application environment on Longleaf is presented as modules using lmod. Please refer to the Help document on modules for information and examples of using module commands on Longleaf.

Modules are essentially software installations for general use on the cluster. Therefore, you will primarily use module commands to add and remove applications from your Longleaf environment as needed for running jobs. It’s recommended that you keep your module environment as sparse as possible.

Applications used by many groups across campus have already been installed and made available on Longleaf. To see the full list of applications currently available run

module avail

Users are able to create their own modules.

Job Submission

Job submission is handled using SLURM. In general there are two ways to submit a job. You can either construct a job submission script or you can use a command-line approach.

Method 1: The Submission Script Create a job submission script using your favorite UNIX editor: emacs, vi, nano, etc. If you don’t have a favorite editor, you can use nano (for now).

nano example.sh

The following script contains job submission options (the #SBATCH lines) followed by the actual application command.

In this example, you would enter the following into your script (Note: that each SBATCH switch below has two ‘-‘ characters, not one):

#!/bin/bash
#SBATCH --ntasks=1
#SBATCH --time=1:00
#SBATCH --mem=100

hostname

In this example, the command you are running is the “hostname” command and the the job submission options request 100MB of memory, 1 core, and a one minute time limit.

To learn more about the many different job submission options feel free to read the man pages on the sbatch command:

man sbatch

Save your file and exit nano. Submit your job using the sbatch command:

sbatch example.sh

The equivalent command-line method would be

sbatch --ntasks=1 --time=1:00 --mem=100 --wrap="hostname"

For application specific examples of each method see this help doc.

For your job submissions, you will need to specify the SBATCH options appropriate for your particular job. The more important SBATCH options are the time limit (––time), the memory limit (––mem), and the number of cpus (––ntasks), and the paritition (–p). There are default SBATCH options in place:

  • The default partition is general.
  • The default time limit is one hour.
  • The default memory limit is 4 GB.
  • The default number of cpus is one.

Longleaf Partitions

Longleaf is set up into different SLURM partitions which you can see by running the command

sinfo

If you’ve read the section on System Information then you are familiar with the different node types available on Longleaf. The different nodes are available through different SLURM partitions:

  • the general compute nodes are suitable for most user jobs and are available in the general partition.
  • the large memory compute nodes are for jobs that need a very large amount of memory are accessible through the **bigmem **partition. Please email research@unc.edu to request access to this partition.
  • the GeForce GTX1080 gpus are accessible via the gpu partition; the Tesla V100-SXM2 gpus are accessible via the volta-gpu partition. See the GPU page for more details on using this resource.
  • The snp (Single Node Parallel) partition is designed to support parallel jobs that are not large enough to warrant multi-node processing on Dogwood, yet require a sufficient percentage of cores/memory from single node to be worthwhile scheduling a full node. Please email research@unc.edu to request access to this partition.
  • the interact partition is for jobs that require an interactive session for running a GUI, debugging code, etc.

To see the count of CPUs in each node allocated to each partition: sinfo -a -N –format=”%N %c %P”

More SLURM Commands

To see the status of all current jobs on Longleaf:

squeue

To see the status of only your submitted jobs on Longleaf:

squeue -u <onyen>

where you’ll need to repalce <onyen> with your actual ONYEN.

To cancel a submitted job:

scancel <jobid>

where you’ll need to replace <jobid> with the job’s ID number which you can get from the squeue command.

To check the details of a completed job:

sacct -j <jobid> --format=User,JobID,MaxRSS,Start,End,Elapsed

where you’ll need to replace <jobid> with the job’s ID number. The items listed in ––format are specified by the user. In this example, we’ll get the user name, job ID, maximum memory used, start time, end time, and elapsed time associated with the job.

For more information on the sacct command

man sacct

To check limits: To ensure fair usage, resource limits (e.g. time, space, # cpus, # jobs) are set in SLURM. System administrators set SLURM resource limits for each partition and qos. In addition, there are limits users and groups (PIs) have [limits are set generally for all users and do not vary by user; same with groups].

  • To see the current limit settings by partition, use the scontrol SLURM command:

    scontrol --oneliner show partition
    or 
    scontrol --oneliner show partition general
    
  • To see the current limit settings by qos, use the sacctmgr SLURM command:

    sacctmgr show qos format=name%15,mintres,grptres,maxtres%20,maxtrespernode,maxtrespu%20,maxjobs,mintres,MaxSubmitJobsPerUser
    
  • To see the current limit settings applied to groups, use the sacctmgr SLURM command:

    sacctmgr show account withassoc where name=<pi_group name of the form: 'rc_' + pi_name + '_pi'>
    

    To find the PI group name assigned to your account, use the rc_*_pi name listed for your onyen in this command: sacctmgr show association user=<onyen>

  • To see the current limit settings applied to users, use the sacctmgr SLURM command:

    sacctmgr show association user=<onyen> format=GrpTRES%40
    

Note: System administrators frequently tweak all settings in SLURM to tailor the cluster to the current workload.

 

Last Update 12/1/2024 3:41:20 PM