FAS Research Computing - Notice history

All systems operational

Status page for the Harvard FAS Research Computing cluster and other resources.

Cluster Utilization (VPN and FASRC login required): Cannon | FASSE


Please scroll down to see details on any Incidents or maintenance notices.
Monthly maintenance occurs on the first Monday of the month (except holidays).

GETTING HELP
https://docs.rc.fas.harvard.edu | https://portal.rc.fas.harvard.edu | Email: rchelp@rc.fas.harvard.edu

The colors shown in the bars below were chosen to increase visibility for color-blind visitors.
For higher contrast, switch to light mode at the bottom of this page if the background is dark and colors are muted.
Operational

SLURM Scheduler - Cannon - Operational

Cannon Compute Cluster (Holyoke) - Operational

Boston Compute Nodes - Operational

GPU nodes (Holyoke) - Operational

seas_compute - Operational

Operational

SLURM Scheduler - FASSE - Operational

FASSE Compute Cluster (Holyoke) - Operational

Operational

Kempner Cluster CPU - Operational

Kempner Cluster GPU - Operational

Operational

Login Nodes - Boston - Operational

Login Nodes - Holyoke - Operational

FASSE login nodes - Operational

Operational

Cannon Open OnDemand/VDI - Operational

FASSE Open OnDemand/VDI - Operational

Operational

Netscratch (Global Scratch) - Operational

Home Directory Storage - Boston - Operational

Tape - (Tier 3) - Operational

Holylabs - Operational

Isilon Storage Holyoke (Tier 1) - Operational

Holystore01 (Tier 0) - Operational

HolyLFS04 (Tier 0) - Operational

HolyLFS05 (Tier 0) - Operational

HolyLFS06 (Tier 0) - Operational

Holyoke Tier 2 NFS (new) - Operational

Holyoke Specialty Storage - Operational

holECS - Operational

Isilon Storage Boston (Tier 1) - Operational

BosLFS02 (Tier 0) - Operational

Boston Tier 2 NFS (new) - Operational

CEPH Storage Boston (Tier 2) - Operational

Boston Specialty Storage - Operational

bosECS - Operational

Samba Cluster - Operational

Globus Data Transfer - Operational

Notice history

Jul 2023

Numerous nodes in bad state
  • Resolved
    Resolved

    The affected partitions are back up within acceptable operating percentages.

  • Identified
    Identified

    Numerous nodes are in a bad state (marked 'drng', but not those marked 'drain'). We are aware and are working to re-open these nodes as quickly as possible.

    This includes but is not limited to nodes in: argelles-delgado, kozinsky, seas_compute, kempner, and other partitions.

Software modules not available
  • Resolved
    Resolved

    Modules should be available again across Cannon and FASSE. If you have an open session, please log out and log in again to use modules.

    This incident has been resolved.

  • Investigating
    Investigating

    An unexpected bug has broken software module availability at this time. Module commands such as module load, module spider, etc may not work.

    We are currently investigating this incident.

Access Issues with holylabs storage server
  • Resolved
    Resolved

    This incident has been resolved.

  • Update
    Update

    We have brought the server back online in a "workaround" state and will be continuing to investigate the root cause. Access should be normal.

  • Investigating
    Investigating

    Lab Storage on holylabs file server is currently inaccessible. We are currently investigating this incident.

boslfs02 down
  • Resolved
    Resolved

    This incident has been resolved.

  • Investigating
    Investigating

    boslfs02 is currently inaccessible. We are investigating this incident.

A100 GPU issue
  • Resolved
    Resolved

    Firmware updates have resolved this issue.

  • Update
    Update

    The proscribed driver update did not fix this issue.

    We are working with Nvidia to find a fix. As a stop-gap nodes which get stuck/flagged will be marked in Slurm for reboot and rebooted once empty of jobs.

    If a A100 GPU host your lab owns is stuck in a bad state, please let us know and we will mark and reboot it.

  • Identified
    Identified

    A100s are open but still experiencing some instability. Infrequently, you may hit the issue we noted earlier.

    We are continuing to work on a solution.

  • Investigating
    Investigating

    An NVIDIA bug may be causing failures in A100 GPUs. nvidia-smi command is slow, or outputs "ERR!" or "No GPUs are found".

    Jobs on affected nodes may fail.

    We are investigating this issue.

Jun 2023

Mounting storage from campus network
  • Resolved
    Resolved

    This incident has been resolved.

  • Investigating
    Investigating

    Due to a HUIT networking issue, mounting storage from wired campus connections may not work.

    We are working with HUIT to resolve this.

    In the meantime, you should still be able to mount storage while connected to the RC VPN.

Emergency maintenance 6/20/23 - Cannon Slurm Scheduler
  • Planned
    June 20, 2023 at 11:00 AM
    Planned
    June 20, 2023 at 11:00 AM

    WHAT
    Emergency Cannon cluster scheduler maintenance Tuesday June 20th from 7am - 11am (actual end time estimated). Running jobs will be paused during maintenance.

    Note: FASSE is unaffected by this.

    WHY
    As most of you know, there are on-going issues with the Cannon scheduler after the downtime and OS upgrades. This behavior did not show up in our earlier testing and only manifested once we were under load from user jobs again.

    After investigation we found that the scheduler is oscillating between two states every hour. During even hours the scheduler would enter a high thread state where all the traffic would pile up, causing any queries to either take a long time or time out. During odd hours the scheduler would enter a low thread state and operate normally. During periods where the scheduler is responding, jobs can be submitted and should then run normally.

    Unfortunately the root cause is still not known, and we are working closely with the vendor, SchedMD, to find a solution to return the cluster to stability. The version of slurm that we are running is 22.05.7, which while very stable — we have been running on it for the past 5 months with no problems — appears to be having issues at scale on Rocky 8.

    There is a newer version of the scheduler available, 23.02.3. This new version has a host of improvements. While the specific issue described above is not listed as one of the items fixed in this version, we anticipate that the upgrade may fix this issue owing to the various code improvements contained therein.

    Barring any solutions found over the long weekend, we intend to hold an emergency maintenance period on Tuesday June 20th starting at 7am. Running jobs will be paused. We will return Cannon to service as soon as we've completed work. Status of this maintenance will be tracked here on our status page.

    We thank you for your patience and understanding. Our team is hard at work answering your tickets and making sure that the cluster is back to full health. Also as a reminder that we hold additional Office Hours each week in June. See: https://www.rc.fas.harvard.edu/training/office-hours/

    FAS Research Computing
    https://www.rc.fas.harvard.edu/
    https://docs.rc.fas.harvard.edu/
    https://status.rc.fas.harvard.edu/

  • Completed
    June 20, 2023 at 1:16 AM
    Completed
    June 20, 2023 at 1:16 AM

    CANCELLED

    Emergency maintenance has has been cancelled. A solution to the scheduler issue has been found and Slurm is once again operating normally.

    We thank you for your patience and understanding.

Slurm scheduler slowness/delays
  • Resolved
    Resolved

    This incident has been resolved.

    A solution to the Slurm scheduler issue has been found and implemented. The scheduler is now operating normally.

    Thank you for your patience and understanding.

  • Update
    Update

    PLEASE SEE EMERGENCY MAINTENANCE NOTICE BELOW
    Emergency maintenance Tuesday June 20th 7am-11am. Running jobs will be paused.

    We are still working on finding the root cause of the scheduler slowness and timeouts. We are working with SchedMD and have sent diagnostic information for analysis.

    At this time the scheduler is up but may become unresponsive to your commands at times. You can wait and retry them.

    Jobs, once launched, should run as expected.

    OOD/VDI are working, but job submission may also be affected.

    If you receive salloc: error: Job submit/allocate failed: Invalid account or account/partition combination specified, waiting a moment and re-submitting your job again should result in a successful submission.

    Addendum: It was determined that the issue is cyclical and you will have the best luck during odd hours (e.g. - after 1pm, before 2pm) and less success interfacing with the scheduler during even hours (e.g. - after 2pm, before 3pm).

    We regret the impact this is having on your work. Updates will follow as we have them.

  • Update
    Update

    We are still working on finding the root cause of the slowness and are working with the vendor to troubleshoot.

    At this time the scheduler is up but may become responsive to your commands at times. Jobs, once launched, should run as expected.

  • Update
    Update

    Slurm is still being overwhelmed with large numbers of requests. VDI timeouts/502 errors, interactive job connections, job submissions, are all impacted.

    We are in contact with the Slurm developers to find a better resolution.

  • Monitoring
    Monitoring

    We have implemented updates and the scheduler appears stable and responsive.
    We will continue to monitor and update this incident.

  • Investigating
    Investigating

    We are currently investigating performance issues with the scheduler. This manifests in slowness when submitting jobs, job submission fails some percentage of the time, and timeout errors (including OOD/VDI).

    Also, some jobs may not exit cleanly putting nodes into drain status.

    We are working to resolve this issue. Updates to follow.

Cannon and FASSE OOD (VDI) issues
  • Resolved
    Resolved

    This incident has been resolved. FASSE and Cannon OOD are both available.

  • Monitoring
    Monitoring

    Cannon and FASSE OOD are not fully operational due to slurm being slow. With slurm slow, the apps wait too long to be launched and it can throw 502 errors

  • Identified
    Identified

    Cannon OOD is operational.

    FASSE OOD is still experiencing issues and showing errors for new connections. Existing connections are working.
    WIP

  • Investigating
    Investigating

    We are currently investigating issues with OOD (OpenOnDemand aka VDI).

fasselogin nodes
  • Resolved
    Resolved

    Login to fasselogin is working. Additional login nodes are not in the rotation and will be added back once they are healthy again. But normal login is working.

  • Investigating
    Investigating

    We are currently investigating an issue with fasse login nodes (fasselogin) timing out or not connectring.

May 2023

No notices reported this month

May 2023 to Jul 2023

Next

For issues not shown here, please contact FASRC via
https://portal.rc.fas.harvard.edu or email rchelp@rc.fas.harvard.edu