Policies and Best Practices

System Usage

Login Node Usage

The login (front end) nodes are a part of the service nodes–providing access to the rest of the cluster. Login nodes are not intended for computation, instead they should be used for code and batch job management tasks. Running heavy processes directly on the login nodes may negatively impact other users who interact with the cluster.

Login nodes should be used for tasks similar to the following:

• Editing and compiling code

• Organizing data on project and home directories

• Submit jobs (batch, dtn, etc, …)

• Monitor jobs

Use compute nodes for processes that require more cores, longer run times, or more memory.

Cron and scrontab usage

To schedule recurring jobs, cron is provided for users. On most systems, cron jobs will be started on service or login nodes; therefore the login node usage policy applies to cron jobs.

However, on systems such as Gaea and PPAN, cron is no longer permitted. Instead, you must use scrontab.

Any process launched from cron or scrontab must be one of the following:

• A call to a workflow manager for controlling recurring tasks

• other scripts that submits or manages jobs in the batch system

• short running (less than 2 minute) data manipulation task

To ensure cron and scrontab jobs do not overload the service or login nodes, please consider the following:

Cron/Scrontab Dos

• Review your crontab/scrontab entries once a month and remove unneeded entries to assure efficient RDHPCS resource utilization! It is easy to forget to remove your crontab entries at the conclusion of your experiment.

• Remove all “busy waits” (tight loops without sleeps) from workflow management scripts.

• Add a line MAILTO=First.Last@noaa.gov to the top of your crontab or add scrontab directive --mail-user=<email> to your scrontab, so you get emails for cron job output (errors).

• Launch additional Slurm jobs using sbatch to the appropriate cluster/partition for compute or long-running data manipulation tasks.

Cron/Scrontab Don’ts

• Do not redirect the output of your cron jobs to /dev/null. This will make it impossible for you to be notified of problems with your cron job.

• Do not rely on cron jobs running on specific hosts. Your crontab will be visible from all front ends, and jobs may be launched from any available frontend. (See Chron for more details).

• Do not run scp, rsync, and tar processes which last more than 2 minutes directly on the cron servers. They are most likely suited for the service or one of the serial parallel environments:

• Do not have a run frequency shorter than every 10 minutes outside an approved real-time reservation, or nor shorter than three (3) minutes when in an approved real-time reservation.

• Do not run compute tasks or large data management tasks from cron, as they will have a negative impact on all users and may cause system instability problems. Such processes may be killed without warning.

Allocations

RDHPCS System compute allocations are determined by the RDHPCS Allocation Committee (AC), with oversight from the NOAA HPC Board. Approved System allocations are typically given to portfolios as a percentage of the System or an average core-hours per month. Each portfolio is represented on the Allocation Committee and an Allocation Committee Chair is assigned by the HPC board typically for a 1+ year term. Each portfolio has a portfolio manager (PfM) who is responsible for managing their projects and Principal Investigators (PIs), and distributing their allocation amongst their projects as needed on each System where they have an allocation. Within a portfolio, allocations on a System can be traded by the PfM as desired. Portfolios may trade allocations with each other on a System or between Systems with approval from all concerned PfM’s, and with documentation and communication with the AC, but this is typically done only for a specified period of time. The PfM (in conjunction with the PI’s) is also responsible for managing disk quota and archive tape usage. A portfolio’s disk quota on a system is initially based on their percentage of compute allocation on that System.

Click to review the On-Premises Allocation Request Form

The Allocation Committee (AC) is appointed by the HPC Board to manage allocations across RDHPCS resources. The Committee has a rotating chair and representation from the Line Offices and Labs which use RDHPCS resources. The AC assigns a monthly compute allocation and approximate maximum disk quota, per RDHPCS resource.

This document lists committee members and PfMs For information on how allocations are implemented on a System, see Priority and Fairshare.

Request an Increase in Allocations

There are three steps to obtain an increase in allocation:

1. Identify your Portfolio Manager (PfM).

2. Request that your PfM complete an On-Premises Allocation Request Form.

3. The PfM completes and submits the request for approval.

4. The PfM opens a Help Ticket to notify RDHPCS of the request. Send email to rdhpcs.<system>.help@noaa.gov, using the actual system name, with Allocation in the subject line.

Adding a Project to an Allocation

Requests for additional project allocation are submitted by the PfM through OTRS or AIM. If the request involves different portfolios, both PfMs will need to approve and accept the transfer. The request should contain the following:

• FROM Project: The project where hours will be subtracted.

• TO Project: The project where hours will be added.

• AMOUNT: The core hours to be moved in the transfer.

• TIME: Is this a temporary or permanent transfer? If temporary, please include the date when you would like this transfer to be reverted.

Note

Allocation increases for a project are constrained by the amount of compute resources designated to a portfolio by the AC. If additional compute is needed beyond the scope of the portfolio’s resources, PfMs may donate or trade hours as desired. Requests for new or increased allocations beyond the allotted portfolio amount on a system should be emailed to the Allocation Committee Chair, as they must be approved by the Allocation Committee.

Cloud Computing Allocations

To request allocation for a new project, complete the Cloud Computing Allocation form. After you complete the form, create a Cloud help ticket to track the issue. Send email to rdhpcs.cloud.help@noaa.gov, using Cloud Allocation Request in the subject line.

Note

The Cloud Fiscal Year cycle starts collecting Cloud HPC requirements in July/August for use in the next FY PoP. If you are requesting an Allocation to a brand-new project outside of the FY cloud Allocation cycle, use the requirements sheet above and follow those directions.

To request an increase in current allocation, submit a Cloud help ticket. Send email to rdhpcs.cloud.help@noaa.gov, using Allocation Increase in the subject line.

Quotas

Requesting Additional Storage for a Project

When requesting additional storage quota, please be mindful of project space usage. Remember that the scratch spaces are not for long term storage. Please utilize HPSS for long term storage.

Submit requests for additional quota via an OTRS help ticket from the PfM. The request should contain the following:

• AMOUNT: The amount of quota needed.<br>

• JUSTIFICATION:The reason why this space is needed.<br>

• TIME FRAME: Is this a temporary or permanent implementation? If temporary, please include the date when you would like this increase to be reverted.

File System Usage Practices and Policies

High Performance File System (HPFS - Scratch)

Hera’s /scratch(1,2), Jet’s /lfs(1,4), and Niagara’s /collab1 are scratch file systems for your input and output project data for running current jobs, NOT for long term data storage.

Warning

Data on scratch is NOT backed up.

1. Keep source code and critical configuration files on /home, and back up critical data to HPSS.

2. Data unused over 30 days is considered old and should be removed or moved to a different storage vehicle.

Note

Use this google doc to assist you with Hera/Jet Scratch File Management

3. Tar up old small files (or delete them) to free up space on the SSD pool and stay under your file count quota.

4. Large files are still optimal for HPC batch job performance.

5. Do not open with O_APPEND unless you really need it.

6. High performance file systems such as Lustre Filesystems are designed for high performance and high resiliency. While they are good at what they do, they’re not very good for very high Meta data intensive operations such as du, find, etc. Please avoid running such commands to monitor your file space space usage, especially from the top of your project level directory.

For more information about projects, see Slurm

General Parallel File System (GPFS)

Gaea’s /gpfs/f5 is a general parallel file system which provides project directories for short term project data. F5 is not backed up. Users must ensure important files are replicated to another off-site location

/data_untrusted

Every RDHPCS user is provided a user directory in the /data-untrusted directory on each HPFS (scratch) file system on RDHPCS systems (Hera, Jet, Niagara, etc.) they have access to.

Your “$SCRATCH/data_untrusted/$USER” directory is provided so that you can move data on and off of the system from any external site, and is for transient data only. Data should be removed from this directory as soon as it’s transferred to its final destination.

Your inaction to remove old data from this directory could negatively impact other users on the system, therefore; failure to comply with this policy will force us to remove your data and disable your access to this directory.

HFS

The /home file system (HFS) is for small amounts of critical labor-intensive data, like source code, that need timely access. The HFS is backed up nightly and weekly. Nightly backups are kept for a week, and weekly backups are kept for at least 6 months.

HFS data can be retrieved from our snapshots - please see Recover Recently Deleted Files from /home for more information.

Each RDHPCS user is given a home directory (/home/First.Last) and a 50GB quota on each system (Hera, Jet, etc.) they have an account on. All files owned by you in /home are counted not just files in your /home/First.Last directory.

Usage and quota can be checked using the sacccount_params or the quota commands. See Slurm for details.

If more quota is required, start a system help ticket with a request and justification.

Caution

Please DO NOT run jobs against files in your Home File System (HFS). This includes keeping input/output files or executable files for a parallel run in your home directory or even using symlinks in your home directories that point to your files in your project space in the scratch filesystem. It puts a tremendous burden on the HFS and has an adverse impact on all the users on the system.

Filesystem Backup and Data Retention

• /home

  • For code and important source files

  • Is backed up nightly. Look at the snapshot directory (/home/.snapshot) to see what options are available

• /scratch

  • For data to drive models and model results

  • Scratch file systems are not backed up

  • Scratch file systems are not purged, it is up to the individual projects to determine the purge policy

  • Only copy new data to the HSMS, it is a scarce resource and is not infinite.

All of the file systems with the exception of /home and the HSMS are considered scratch file systems. We want users to be aware of the risks with those file systems. The scratch file systems are not backed up in any way. While we do our best to ensure the reliability of those file systems, they are not perfect and there is a small risk that data could be lost without any possibility of recovery.

Recover Recently Deleted Files from /home

Differences between the HPCS

The home filesystem is backed up regularly. However, the filesystem also supports snapshots, which will allow you to retrieve your own files if they have been deleted over the last few days. The number of days is different for Hera and Jet clusters.

Look at the snapshot directory (/home/$USER/.snapshot) to see what options are available. Each directory listed there represent a day.

Consider an example:

$ ls $HOME/.snapshot
2021-09-09_0015-0600.daily  2021-09-14_0015-0600.daily  2021-09-19_0015-0600.daily
2021-09-10_0015-0600.daily  2021-09-15_0015-0600.daily  2021-09-20_0015-0600.daily
2021-09-11_0015-0600.daily  2021-09-16_0015-0600.daily  2021-09-21_0015-0600.daily
2021-09-12_0015-0600.daily  2021-09-17_0015-0600.daily  2021-09-22_0015-0600.daily
2021-09-13_0015-0600.daily  2021-09-18_0015-0600.daily  2021-09-23_0015-0600.daily

You can then access the old files in your copy of your home directory under the appropriate snapshot.

So, if you want to recover files in your $HOME from January 22nd, 2024:

$ cd $HOME/.snapshot/2021-09-22_0015+0000.homeSnap

Copy the files you want from the here, the snapshot, to anywhere in your real home.

HPSS (Data Retention)

Retention based storage is the HPSS archive policy in Fairmont, to better manage data growth.

Six retention storage pools (1-5year and Permanent) were created. Each retention period is set up as a separate file family. This means all data for a given retention period is stored on the same tapes.

All HPSS projects were then configured to write to one or more of these pools. Data in these pools expires based upon the retention pool it was written in and would be deleted upon expiration.

All files in the HPSS archive have been assigned an expiration date based on the file create time and the retention period it was written to. Upon expiration files will be deleted from the HPSS archive.

Expired Data Deletion Process

User Notification

Users will be notified of expired data via posted lists and email. These notifications will take place on or before the first day of the month following the data’s expiration.

For example, data that has an expiration date between October 1 and October 31 2023 will have its notification posted on or before November 1, 2023. The expired file list is located on HPSS in /Expired_Data_Lists/expired.YYYY-MM.txt. All HPSS users have read access to this file and can retrieve it for review. The file is easily searchable by HPSS username.

For each file included in the expired list the file owner, file group, filename/path, and expire date are shown, for example:

root system /1year/SYSADMIN/nesccmgmt/test_file-1G-11 Jul-6-2023.

Email notification will also be sent to all users who have data listed in this file. It is the user’s responsibility to regularly check the posted list for expired files they own. Once deleted these files cannot be recovered.

Expired Data - Deletions

The following table maps out when future deletions will take place.

Expire Date	Notification Date	Delete Date
Dec 1 – Dec 31	January 1	February 1
Feb 1 – Feb 28	March 1	April 1
Mar 1 – Mar 31	April 1	May 1
Apr 1 – Apr 30	May 1	June 1
May 1 – May 31	June 1	July 1
Jun 1 – June 30	July 1	August 1
Jul 1 – Jul 31	August 1	September 1
Aug 1 – Aug 30	September 1	October 1
Sept 1 – Sept 30	October 1	November 1
Oct 1 – Oct 31	November 1	December 1
Nov 1 – Nov 30	December 1	January 1

Data Recovery Policy

Occasionally an archive tape is damaged or otherwise becomes partially unreadable. When that happens, the local RDHPCS staff works with the manufacturer to troubleshoot the problem and take additional steps to attempt to recover the missing data.

Very rarely, even with these additional efforts, we are unable to recover missing files. The user will be told which files we cannot recover.

In that case, the user has one further option. There are a number of outside recovery services which will make further attempts at recovery for a fee. Some charge a flat fee, some charge more if they are able to recover than if they are unable to recover.

If the user wishes to sign up for such a service and pay the fee, RDHPCS will handle the logistics of shipping and other coordination with the recovery service.

Data Disposition

RDHPCS users’ data is the responsibility of the user, the PI, and the Portfolio Manager. The PI or Portfolio Manager, as appropriate, can initiate a help request to manage data. As a policy matter, RDHPCS System Management does not initiate the deletion of data belonging to active users or active projects, except as detailed below.

HPFS (Scratch) Data

Inactive users’ and closed projects’ data shall be dispositioned by the PI or Portfolio Manager to maintain efficient usage of RDHPCS resources. If the PI or Portfolio Manager cannot personally implement the disposition of the data, the PI or PM can issue a help ticket, and request that RDHPCS System Management do so.

The RDHPCS program policy is to NOT delete active project HPFS data. If the PI or Portfolio Manager so directs in a help request, we will change ownership of active HPFS project data to another project member.

Niagara Per User Data

As Niagara is a hybrid system (a cross between a traditional HPC system and a data transfer/collaboration system, available to all RDHPCS users), the file system management needs to be handled differently then on more traditional HPC systems (Hera and Jet). As a result, the following data management policies are implemented on Niagara:

• All files under the collab1/data_untrusted/$USER directory tree which have not been accessed in the last 5 days will be automatically purged.

• All files under the /collab1/data/$USER directory tree which have not been accessed in the last 60 days will be automatically purged.

• All files under the /collab1/data/$PROJECT directory are treated the same as HPFS (scracth) data and are not deleted.

The definition of access time is the last time the file was opened for reading or writing.

Note

If the file system’s usage starts getting close to the total capacity, we will be forced implement a more aggressive purge policy (i.e. 30 day or 15 day purge) . So please actively manage your data.

Home File System (HFS) Data

The RDHPCS program policy is to NOT delete active users Home File System (HFS or /home) data, or to change ownership of HFS data. The Portfolio Manager may issue a help ticket to request special dispositioning of HFS data.

Deactivated users’ HFS data may be removed and saved to the tape archive system in a retention pool of at least 5 years.

Protecting Restricted Data

This describes the process to protect the RSTPROD restricted data on Hera. Hera uses regular Linux group based protection for restricted data.

It is up to the user to make sure that files containing restricted data are set to have the group as rstprod and also to make sure that permissions for the world are removed.

# chgrp -R rstprod $DIR
# chmod -R rwx-go $DIR

Where $DIR is the directory with the files you want to protect.

When these files are copied to a different location, be sure to use the -p option on the cp command, to preserve the group and the protection for those files:

# cp -rp $DIR $TARGET_DIR

Managing Packages in /contrib

Overview of contrib Packages

The system staff do not have the resources to maintain every piece of software requested. There are also cases where developers of the software are the system users, and putting a layer in between them and the rest of the system users is inefficient. To support these needs, we have developed a /contrib package process. A /contrib package is one that is maintained by a user on the system. The system staff are not responsible for the use or maintenance of these packages.

Responsibilities of a contrib Package Maintainer

Maintainers are expected to:

• Follow the naming conventions and guidelines outlined in this document

• Apply security updates as quickly as possible after they become availble

• Update software for bug fixes and functionality as users request

• Respond to user email requests for help using the software

contrib Packages Guidelines

• The package should be a single program or toolset. We want to prevent having a single directory being a repository for many different packages.

• If you support multiple functions, please request multiple packages.

• The package may have build dependencies on other packages, but it must otherwise be self-contained.

• The package may not contain links to files in user or project directories.

• We expect each package to be less than 100MB.

• If you need more, please tell us when you request your package.

• We can support larger packages but we need to monitor the space used.

• We expect each package to have less than 100 files.

contrib Package Maintainer Requests

If you wish to maintain a package in contrib, please send a request to the Help System with:

• List of the packages you wish to maintain.

• Justification why each is needed.

• The user who will be maintaining the package.

Note

In certain cases, multiple users can manage a package, and unix group write permissions may be granted for the directory. In that case, specify the unix group that will be maintaining the package.

Managing a contrib Package

After your request has been approved to use space in the /contrib directory, two directories will be created for you:

• /contrib/<package>, and

• /contrib/<package>/modulefiles

This is where you will install your software for this package and optionally install a module to allow users to load the environmental settings necessary to use this package. The variable <package> is the name of the /contrib package you requested. The directory convention of /contrib is designed to match that of /apps. Thus, one piece of software goes into a subdirectory under the /contrib level. If you want to manage multiple packages, please request multiple /contrib package. You can do this all at one time when submitting your request to the Help System.

Maintaining “Metadata” for contrib Packages

Since contrib packages are intended to be used by other users on the system it will be helpful to have a /contrib/<package>/README file that contains at least the following information:

• Package Name:

• Purpose:

• Maintainer:

• Contact info for questions/help:

• Any other info that will be useful for general users to know

contrib Package Directory Naming Conventions

When installing software into your /contrib directory, first determine if this is software that should be versioned (multiple versions may exist at one time) or unversioned (there will only ever be one version installed, and upgrade will overwrite the existing software). For verisoned software, please install it into a subdirectory of your package that is named after the version number. For supporting multiple versions of software the install path should be:

/contrib/<package>/<version>

Where <package> is the directory assigned to you and $VER is the version number. Thus if your package is named ferret and you are installing the version 3.2.6, the software should be installed in:

/contrib/ferret/3.2.6

For supporting un-versioned software, only install the software directly into your package directory:

/contrib/<package>/

Queue Policy

Overview

• The queuing system should allow groups/projects to spend their allocation each month.

• The tension between keeping persistent jobs in the system and running very large jobs suggests that there should be a limit on the number of cores a job may use, but with a capability to make exceptions for “novel” jobs that may require up to the entire system.

  This will promote consideration of whether a job requires a large number of cores due to, for example, memory or schedule constraints, or whether it is simply desired.

• There should be queues with different priority levels usable by the scheduling algorithm. At the very least, run-time variability would need to be assessed before we could even think of implementing this.

Specifying a Quality of Service (QOS)

To specify a quality-of-service (QOS), use –qos (-q).

For example, to specify the batch QOS:

$  #SBATCH -q batch

Several different QOS’s are usually available.

Changing QOS’s

You can change the QOS of jobs at submission and post submission. While you can use this feature in many different ways, one practical situation where this may be useful is to maintain your fairshare priority by starting jobs in the “windfall” QOS, then changing to the “batch” QOS if it is still pending. See Slurm_ for more information on Fairshare

Note

If your job does not meet the criteria of the QOS that you change it to, it will remain pending indefinitely.

You can immediately change the QOS of your pending job(s).

The following is an example of immediately changing 2 pending jobs (26866 and 26867) to the “batch” QOS:

$ scontrol update job 26866,26867 qos=batch

When submitting a job to a certain QOS, you can tell Slurm to change it to a different QOS at a certain time if it is still pending. In the following example, you submit the job to the “windfall” QOS, then tell Slurm to change the job to the “batch” QOS if it’s still pending after 5 minutes. NOTE: Do not use a time less than 2 min (120 seconds).

Note

On Orion and Hercules the “at” functionality is only available on login1.

$ sbatch -q windfall jobfile
Submitted batch job 26990

$ echo scontrol update job 26990 qos=batch | at -M now +5min
warning: commands will be executed using /bin/sh
job 6 at Sun Dec 17 16:07:00 2023

You can change the QOS of all your pending job(s) in a QOS to another QOS after it has been pending for a certain time. The following example script will change all your pending “windfall” jobs to “batch” if they have been pending for at least 600 seconds (10 min), whenever you run it.

Note

Do not use a time less than 120 seconds (2 min).

Jet and Hera

Note

If you have an allocation of “windfall only” (Allocation = 1) you can only submit to the windfall or gpuwf QOS.

QOS	Minimum Nodes	Maximum Nodes	Maximum Wall Clock	Billing TRES Factor	Description and Limits
All QOS’s					Max of 400 pending/running jobs per project/account, additional jobs will be rejected. Max of 20 jobs per project/account will gain age priority. Exceptions are stated below.
batch	1	8,400[1]	8 hours	1.0	Default QOS for non-reservation jobs with an allocation more then Windfall-Only (RawShare=1).
urgent	1	8,400[1]	8 hours	2.0	QOS for a job that requires more urgency than batch. Your project’s FairShare will be lowered at 2.0x the rate as compared to batch. Only one job per project/account can be pending/runnin at any time. When a project’s FairShare is below 0.45, jobs submmit to urgent are automatically changed to batch and users notified via stderr.
debug	1	8,400[1]	30 minutes	1.25	Highest priority QOS, useful for debugging sessions. Your project FairShare will be lowered at 1.25x the rate as compared to batch. Only two jobs per user can be pending/running at any time. This QOS should NOT be used for fast-turnaround of general work. While the debug QOS is available, we recommend that if you need to work through an iterative process to debug a code, that you submit a longer running interactive job to the default QOS so that you can restart your application over and over again without having to start a new batch job.
gpu	20 (1 node)	800 (40 nodes)[1]	168 hours (7 days)	1.0	This QOS can only be used on Hera in combination with the fge partition. Max total “GrpTRESRunMins” of 720,000 core-minutes (600 node-hours) of running jobs at any time, per project-account. “GrpTRESRunMins” is defined as cores_allocated * wallclock_requested of running jobs. A project can have up to the max number of jobs pending/running as defined above, but the queued jobs will NOT be considered for scheduling if the project’s running jobs exceed this limit. Use this gsheet as a reference: Grp TRES Run Minutes For example, the following combinations of the max running jobs per project-account are permitted: 1. A project can have three 1-node jobs at 168 hours of wallclock and one 1-node job at 96 hours of wallclock. 2. A project can have one 8-node job at 75 hours of wallclock.
gpuwf	20 (1 node)	800 (40 nodes)[1]	168 hours (7 days)	1.0	This QOS can only be used on Hera in combination with the fge partition. Max total “GrpTRESRunMins” of 201,600 core-minutes (168 node-hours) of running jobs at any time, per project-account. “GrpTRESRunMins” is defined as cores_allocated * wallclock_requested of running jobs. A project can have up to the max number of jobs pending/running as defined above, but the queued jobs will NOT be considered for scheduling if the project’s running jobs exceed this limit. Use this gsheet as a reference: Grp TRES Run Minutes For example the following are combinations of the max running jobs per project-account that are permitted: 1. A project can have two 2-node jobs at 24 hours of wallclock and one 1-node job at 72 hours of wallclock. 1. A project can have one 1-node job at 168 hours of wallclock. Lowest priority QOS for use with GPU nodes. If you have an allocation of “windfall only” (Monthly allocation = 1) you can only submit to this QOS. Submitting to this QOS will NOT affect your future job priority FairShare Factor (f). EffectvUsage = 0. See How FairShare Works. This QOS is useful for low priority jobs that will only run when the system (partition(s)) has enough unused space available, while not lowering the project’s FairShare priority.
windfall	1	8,400[1]	8 hours (Partition exception: service)	0.0	Lowest priority QOS. If you have an allocation of windfall-only (monthly allocation is 1) you can only submit to this QOS. Submitting to this QOS will NOT affect your future job priority FairShare factor (f) for your non-windfall jobs. Useful for low priority jobs that will only run when the system/partition has enough unused space available while not effecting the project’s FairShare priority.
novel	501 (Orion), 251 (Hercules)	Largest partition size	8 hours	1.0	QOS for running novel or experimental where nearly the full system is required. If you need to use the novel QOS, please submit a ticket to the help system and tell us what you want to do. We will normally have to arrange for some time for the job to go through, and we would like to plan the process with you.

[1] (1,2,3,4,5,6)

Some partitions are small than the Max Cores QOS limit. Jobs submitted only to partitions with an insufficient number of cores will get stuck in pending, will not run, and will have to be manually deleted by the user. The max nodes allowed per partition is the min of the max cores allowed divided by the cores per node of the partition (Hera and kJet: 8400/40=210 nodes) or the max number of nodes in the partition (vJet: 288 nodes).

Gaea

This section documents the queue structure on Gaea. The original queue policy was approved through NOAA’s HPC Integrated Management Team. Changes and fine-tuning to the queue structure can be done on a weekly basis through the Configuration Management process.

The following guidelines were put in place:

General Recommendations

• Use a fair-share algorithm that can throttle scheduling priority by comparing how much of a particular allocation has been used at a given time with how much should have been used, assuming constant proportional usage. This will promote steady usage throughout the month.

• Use two separate allocations, renewed monthly, with multiple queues drawing down each of them:

  • 50% of the available time for high-priority and urgent work. That should minimize queue wait time. Queues are:

    • Urgent, for schedule-driven work that must be completed ASAP.

    • Novel, for jobs that have unusual resource requirements, typically needing more than 25% of the system’s cores. These can be run during an 8-hour period immediately after Preventative Maintenance is complete, since no other jobs will be running at that time.

  • 50% for all other normal-priority allocated work. Queues would be:

    • Batch, for regular allocated jobs

    • Debugging/Interactive work

    • Windfall, a quality of service (QOS) tag, for work that will not be charged against an allocation. Windfall can be specified with ‘-l qos=’ directive, as:

$ sbatch –-qos=windfall

or in your job script:

#SBATCH -–qos=windfall

Priorities Between QOS

• Normally, the Urgent QOS will have the highest priority but remain subject to the fair-share algorithm. This will discourage groups from hoarding high-priority time for the end of the month.

• Within a group/project, jobs in the Urgent queue are higher priority than jobs in the Normal queue, with each group expected to manage the intra-group mix per their allocation.

• At any given time, the suite of jobs drawn from the Urgent queue and running on the system should use about 50% of the available cores (per the fair-share algorithm), but that suite is permitted to use more than 50% as needed (with the implication that less than 50% will be used at other times of the month).

• Limit the largest job to 25% of the available cores except in the Novel queue.

• Limit time requested for individual job segments to 12 hours.

• Interactive/debugging jobs have a tiered limit.

Debug & Batch QOS

Interactive / Debug The interactive queue may have different time limits based on the size of the submitted job. To see the current queue wallclock limits, run

$ sacctmgr show qos format=Name,MaxWall